I was on the fence on jumping into the Anthropic IPO - but i think i'm going to skip it even though it's a product I actively pay to use and am happy with. Its just too overpriced.
Yes Anthropic will be leading for a while with the best model; but first the Chinese models, yesterday's OpenAI's Sol & Astra, and now even Microsoft getting Opus 5 performance. Remember that even Anthropic's own Opus 5.5 is _way_ more efficient than the models that were active in their “profitable” quarters.
I do not see a business that is going to do triple digits growth consistently to justify their valuation.
It is just too much of a fungible utility — and there is too much demand for any single approach to capture a market. Hell, AI is the perfect tool for any customer to 'uncapture' themselves.
It's increasingly looking like AI companies have invented electric motors. Incredibly valuable technology that will be in everything. But nobody ever built a moat around electric motors. At least as to LLMs, AI is a competitive, efficient market where the basic interface is the most plug-and-play, standardized thing imaginable: a text stream. Literally anything that can produce a 9600 bps serial TTY stream is basically interchangeable with Claude, GPT, etc.
The essay "AI will not make you rich" by Jerry Neumann makes a similar point, but uses containers (the shipping kind) as an analogy.
It's well argued and the arguments are worth considering.
In the long term, sure. But in the short term, it seems like it will make a few people very rich.
Electric motors have few dimensions compared to models.
I think a more apt analogy might be cars. In each segment you have room for more than one company. The prices are roughly the same, so is the quality, but there can be no car that satisfies everyone. There are different „feels” to them.
Neither analogy really works, because of how easily you can switch between models, and there being no "purchase" cost. No one is switching between five different cars a day...,
Probably too dated but I would say ISPs. In the dial up days a trivial payment gave you access to a literal miracle. And the was enough differences in price and performance it really mattered who you chose. But the right choice varied over time, initially changing was easy, but they got good at locking in and suppressing competition; then overnight didn't matter at all.
ISPs in the US still have a physical moat today around their wires. Most places have one, maybe two viable options for high speed internet.
OP said “in the dial up days”. Back then all you needed was a phone line and a number to dial.
I think AI models will end up like Linux VPSes: a standard commodity available from dozens of providers at a variety of price points.
Maybe not five a day, but those https://www.flexcar.com/ people are inundating me with ads on how I can change my car every month (or something; I have no idea how it actuality works).
The analogy is still off even with Flex Car, but not as far.
You’re right the fewer dimensions makes the analogy imperfect.
But this is more about the electric motor market than the electric motor itself maybe.
A transformative technology that I think the foundation labs expected would be a bit easier to build a moat around or hold competitive advantage.
Sure, you can have differentiated products, but it looks like the moat is small. We can expect monopolistic competition. And monopolistic competition predicts normal profit in the long run.
There can be excess profit for awhile but it seems like there's not a sufficient barrier to entry to prevent an eventual race to the bottom.
Not another car analogy please ...
The teletypewriter truly was the greatest invention in human history.
I just finished "The Victorian Internet". The author made a similar observation that the internet might have looked less impressive to someone teleported from the late 19th vs. the mid 20th century.
Because it would basically appear to be a completely predictable development from the continuous incremental improvements made to the telegraph, which had already eliminated the operator and evolved into mechanical proto-computers.
Alexander Graham Bell stumbled on his discovery while working on harmonic telegraphy (at the time the latest hype cycle in new telegraph tech). Inventors were scrambling to find every possible way to further automate and expand telegraph use cases, so if the telephone hadn't been invented, might have led to increasingly complex mechanical computers linked by the telegraph network.
But instead, the telephone dominated the zeitgeist and text interchange suddenly looked like a soon-to-be obsolete stepping stone towards the telephone.
Victorian England had mail delivery to their house a few times a day. So your could send a dinner invite across town and get a response in time to set a place setting for them.
If you're willing to pay for it there are plenty of courier services able to do this today depending on your locale.
I'm still partial to people having pneumatic tubes running their documents across town for them.
Nah; social engineering via shared language wins
Reinforcement of oral and written traditions helped get people to sit still rather than hunt and gather in loose clusters
That said we also went off the deep end with the allegory and lost sight this is generally about species survival not personal acquisition of stuff; rarely has it gone well when the majority are holding the bag in deference to the few.
Feels like due to modern logistics a regular cadence is possible between nomadic (both in terms of bodily agency and free thought) and dogmatic tribal custom/long term nesting.
Dump the Gregorian calendar, rely on physical metrics to track seasons, and live far more async routines yet stick to share desire to keep logistics flowing
Anyway off topic; climbed down this rabbit hole with others recently and coincidentally saw your comment
Compare electric motors to search engines or operating systems, and your analogy will not work for comparisons to Google and Microsoft/Apple respectively. Computing development costs are fairly different from a max triple digit $ BoM predictable hardware assembly in the form of an electric motor.
That may have been true if one or 2 AI companies maintained an effective monopoly on models and it took other players 10+ years to catch up.
Google maintained an effective monopoly on decent search results for most of its existence. The closest second competitor (who isn’t just licensing their results) was bing, and you could still see a “bing vs google” screenshots of how laughably bad bing was up until 2020.
Similar problem with OS market. No one other than Microsoft and Apple (then later Google) had a lead measured in decades not months.
The relevant difference is barriers to competition. Operating systems have compatibility as a major barrier to competition. Search has network effects--more users = more information for making search results more accurate. LLM models have no such barriers.
> LLM models have no such barriers.
compute and amount of training data built on top of XXX B investments is very strong barrier.
For frontier LLMs? Sure. A almost prehistoric datapoint by now: Training costs for GPT-2 levels of "intelligence" fell by 600x over 7 years, as measured by Karpathy in Jan 2026, source https://x.com/karpathy/status/2017703360393318587
If we can extrapolate from that (I don't know if we can), good enough models might be achievable with much less than XXX B investments.
I don't think we can extrapolate evolved complexity. I think frontier labs have many thousands researchers crafting and testing various post-training tuning datasets with impact on end quality, which was not the case in gpt-2 time.
>LLM models have no such barriers.
I think it's much more like cloud, where sure, there aren't any single barriers to buying some servers, putting K8s on it, having an account management/billing system...but in practice the expertise in aggregate to run a profitable one is quite rare.
I don't know about that. The comparison may make more sense than it seems at first. The whole "war of the currents" was about electric motors at the end of the day, at both the generation and consumption endpoints.
You had a couple of 900-pound gorillas -- Edison (the proto-General Electric, merged by the same JP Morgan whose name is still on the bank) and Westinghouse -- playing the part of OpenAI and Anthropic. Both very valuable long-term players throughout the whole twentieth century.
And both with patent moats a mile wide and deep. We haven't seen that side of things yet, but we will.
Given the progress we have seen in local AI as well this year its quite possible that the bulk of programming needs might soon be entirely met by models that run in a 16GB consumer card as well. They are already remarkably capable of implementing specs and working on PHD level problems at a fraction of the size and cost.
We aren’t going to stop at PhD level work and say “that’s good enough”. If a PhD can run on a consumer laptop, it’s going to look dumb compared to what’s available in a datacenter, and won’t be capable of whatever superhuman work would be the norm in that scenario
PhD levels? Idk, at least in physics and comp sci, before Fable, I wouldn’t call models capable of doing PhD level work.
The models do seem capable of reaching graduate levels. But its very spiky, and typically requires using subagents correctly. Single agent will not usually do it.
I have the same sentiment - while I enjoy using Opus 5.5 and Anthropic models, local AI has seen significant strides. Especially with models like Qwen 4 in the horizon, I could see regular users able to run capable models on 16GB GPUs in the next 1-2 years or even sooner.
This is a viewpoint I've consistently heard and while it's a small sample, I wonder if many folks are silently (or not so silently) thinking it. What is Claude's real moat?
Claude the model has no moat.
Claude the product has a moat.
People build up skills, memories, dashboards, and now documents and wikis in it. It’s getting increasingly sticky in enterprise. They know the model has no moat; now they are using the model to capture corporate workflows that become more difficult to swap out.
It’s pretty easy to tell codex to import and adapt all your Claude skills and memories and vice versa, so I’m not sure how much of a moat all that is until they start making it inaccessible to you.
>I’m not sure how much of a moat all that is
The way I see it is that AI usage is starting to go the way of automobile favoritism. Does it really matter whether someone gets a Ford F-150, a Chevrolet Silverado, or a Dodge Ram? Not really, they're all trucks that can be had in a variety of colors, cab sizes, bed sizes, and engine choices, but you'll find that people can become fanboys of one brand or another for whatever reason, and they generally don't switch unless they experience some colossal failure. I think the same thing is going to happen with AI providers, the favoritism itself is the moat.
I would say its almost like Apple vs Android. Yes, Android could make a phone that could do exactly the same thing. But Apple became "the standard" (at least in America).
Claude is currently doing this -- it is so ubiquitous in corporations that people are building full Claude workflows and harnesses.
Maybe it will be easy to transfer all that to a different LLM, but I think it's also possible that they could end up with a similar Walled Garden as Apple without having a far superior product, per se
Yeah they seem to be chasing a classic integrations and sticky-identity playbook. But it is weird this time because the product itself can trivially recreate or bypass the moat. Phones did not happily and actively unravel their makers’ competitive differentiation when asked to do so by their users.
I think many recognized a while back that their safety and doomerism was (at least partially) a way of trying to create one by baiting governments into regulations they could shape to crystallize their position alongside OpenAI and FAANG.
This is why they are all racing to get app features out to capture people into their moat. In particular, ChatGPT sites and Anthropic Artefacts are both huge traps for businesses that will create a long sticky tail of large subscriptions.
Models have no moat, so they are reverting to the tried and tested enterprise SaaS playbook.
Agreed. What do you think is the path for app developers?
Do we all integrate with chatGPT/Anthropic via MCP and basically become providers of context and data?
Or do we have our own app interface, like SaaS always has had, but with intelligence tightly coupled for the use cases the app is designed for?
But why not buy the initial wave then ditch when (or if) things start to go south?
Because even if that'll happen one day, time is still on their side at least on the mid-term?
> But why not buy the initial wave then ditch when (or if) things start to go south?
Buy the hype, sell when it crashes? Is that what you're suggesting?
Cut him some slack, he used up all his tokens on Fable 5.2 Max.
And a new Gemini model dropped since your comment lol
Sounds like this model release, right before the IPO, had its intended effect.
I bet they release some new models on the day of the IPO
>I was on the fence on jumping into the Anthropic IPO
never do that
Being on the fence seems perfectly safe to me.
much safer to be on the safe side of the fence with nowhere to fall.
I just want to learn, code on my own. I would literally take a 75% pay cut to go work at a place where llms are banned and I can use my own mind, preferably without going into defense. But I dont even care that much at this point, as long as my brain doesnt atrophy and my existance isnt being a meat proxy between the llms and management.
It feels like all the worlds most powerful companies are pushing these things on us to make us stupid and reliant. This timeline sucks so much.
We actually did significant research before llms. You cannot convince me that research and progress is getting done now, outside of like 10 companies.
Every is just using llms to cheat at their jobs and avoid tackling hard problems. I fully expect there to be a complete plateau in human progress because of these things, while its dressed up as the opposite.
Everybody is cheating on their jobs and while thinking to themselves "I bet theres some mega smart people out there using these llms to cure disease" while those super smart people are also just being lazy, cheating and letting their brains atrophy and also hoping someone else is taking the hard stuff seriously.
Gg idiocracy is coming.
Start a business and you can. My income dropped 99% after doing this.
Exactly, anyone can choose to do what they value in life. We call people who do this artists or entrepreneurs.
Noone has to do work they dislike they just have to live with the consequences of living outside of society.
I don’t understand the end of this sentence.
The only guess I have is "do" = "though".
EDIT: Oh they removed "do" entirely ... now we'll never know!
You can't get fired for buying IBM. And when you use AI you're not really responsible for the outputs, AI is. You don't get credit but you cannot get much blame either. Mediocricity among people is on the rise?
>You don't get credit but you cannot get much blame either. Mediocricity among people is on the rise?
You shouldn't get much blame but you absolutely will get the blame from some, the people assigning blame when their decisions go tits up are rarely rational.
This is the classic "Why didn't you tell me? I did, You should have told me harder" or "I know we gave you no resources to fix this even though you asked, now it's broken and that is clearly you fault".
You may have been lucky enough to have had reasonable/competent management in your career but that very much isn't always the case.
Good leaders own their decisions good and bad, so if they force you to use AI and it goes sideways, they should own that, Bad leaders will do the opposite, it'll still be your fault and there is fuck all you can really do about it because of the fundamental power imbalance.
I wonder if it's structurally similar to how smartphones killed pop music.
Up until the early 2010s you'd hear music made within the past 0-5 years and now it seems to all be 15-30 years old.
The dispersion of the tools of engagement and creation has sacrificed the collective. Maybe this is the interregnum or maybe it's the new stagnation.
I guess the other place we can look is how machines have replaced human physical labor and analogize things like "fewer amputees but more cardiovascular disease" or how the decrease in the utility of human labor is related to the rise of modern sport.
>I wonder if it's structurally similar to how smartphones killed pop music.
>Up until the early 2010s you'd hear music made within the past 0-5 years and now it seems to all be 15-30 years old.
---
There is plenty of music being made today? Very confused what you're saying here
If you walk by a bar, turn on FM radio, go into a grocery store, turn on a television, watch a movie, the median age of the song has increased
https://www.yahoo.com/entertainment/music/articles/why-chart...
"Eighteen of the year's top 20 bestselling songs were released 12 months ago, or earlier.
The year's eighth biggest hit is Zara Larsson's Lush Life, which originally came out in June 2015. Michael Jackson's Billie Jean (1983) is in 22nd place. Fleetwood Mac's Dreams (1977) is 30th."
Billboard has also covered this extensively but it's paywalled.
This is also called the decline of the monoculture and has affected profit margins of movies and television shows:
https://www.forbes.com/councils/forbestechcouncil/2025/06/04...
The music claim is specifically of the median age of commercial pop music increasing, the velocity of new celebrity musicians decreasing and the longevity of commercial play increasing.
It's not that fewer things are being put out on bandcamp or less music is being put on youtube.
If you insist that say Ambient Trip Commander by Legowelt is in the same category as 21 by Adele, I will concede there's no way to do this analysis. Of course the derivative of the per annum quantity of catalog releases on discogs is positive.
> If you walk by a bar, turn on FM radio, go into a grocery store, turn on a television, watch a movie, the median age of the song has increased
There are a few things going on here. The increasing age of the music you hear is partly due to the increasing distance we are from the dawn of music recording and proliferation. There's going to be a natural elongation of the timeline for music you hear simply because there's more of it as time goes by.
Now halfway to your point, the frequency of music releases has greatly increased with the lower barrier for high-quality music production. But contrary to your point, there's still a huge barrier for being noticed, and a lot of new artist promotion is still happening through some pretty familiar channels like television and radio. Social media dilutes this somewhat, but new artists often still rely on big money promotion regardless. So in that way you still have a bit of a monoculture.
But music streaming definitely works against the monoculture. There's just so much of it and it's much easier, seemingly, to stumble upon smaller artists.
Pop music for sure is mostly dead. It tries to appeal to everyone while not being particularly enjoyable to anyone. If you look beyond supermarket radio music things are still looking healthy. I'm starting to observe retail store music make the jump from 2014 pop directly to modern house music.
If you are choosing to do the same work as before then you will atrophy. Your job changed and you may not like the new job but so what? Now you should be managing 4-8 ai coding sessions at once and that will push you to your limit and stop the atrophy. But its a different skillset.
If a software engineer became a project manager or lead engineer would you say he is being lazy? Because now he barely writes any code. He just makes decisions about whats important and fixes things occasionally.
People who choose to do the same work as before are who is becoming lazy. Once companies catch up those people will have to adapt or will be fired. If they are kept it will be out of the goodwill of the company.
You can take a 100% paycut at any time to go do stuff that no one currently values. You can become an artist or an entrepreneur and if your ideas are eventually valued by society then you will be rewarded. Noone is stopping you from taking that risk.
Your argument is so lazy about people using llms to cure disease. Idk if this whole post is satire or a joke. If you give an AI agent a large task is will eventually get to a place where it is stuck or constrained by resources. And humans are what can help it get unstuck.
I think the problem is you are attaching your self worth to your ability to write good code and that is making you unhappy as people are lowering that skills value.
The old world valued making software that had few really good features. The new world values making software that has TONS of very good features that all works well.
Actual loser mentality here. The biggest development in the history of technology is happening and your response is literally to want to hide or rewind time back to nonexistent halcyon days. You would have been one of the people hounding Copernicus for heresy.
Progress marches on whether you like it or not. Roll with the punches lest ye be rolled by them. But don’t ask for a soft padded box free from aspects of reality that displease you.
Love your personal website. I agree with your views
Which personal website? I'm interested
Hell yeah. let's ban compilers too. never trust those things.
I didn't watch the keynote, but I've been testing MAI Code 2 and its been quite good for my usual work. They put it on par with Opus or Sol, and I wouldn't argue too much with that. Kudos to them for starting from scratch and getting it right.
> Kudos to them for starting from scratch and getting it right.
Did they? I thought OpenAI had quite bad deal with them and wanted to get away from it. And this was one of the reasons. Microsoft got access to everything.
Right, the “bad deal” where Microsoft gave $13 billion cash and $100B worth of free hosting to a startup that was pre-revenue...
Microsoft invested a lot, but they can use any novel OpenAI ever develops until 2032, and they have much more money to take any advantage of it.
Hmm so the obvious question is are these models just distillations?
I wonder if that contract forbids MS from using openAI models to develop their own.
How do you test it ? Is it on your Azure Foundry?
They will also soon be releasing on OpenRouter, Vercel, Hermes, OpenCode, and Kilo.
Visual Studio's built-in Github Copilot. Though they have a desktop app of their own now.
I thought they were focused on a VS Code extension. Thanks for the info.
For me, GitHub Copilot via VS Code
Is it lower in price then GLM 5.3 Flash and DeepSeek v4.1 Flash?
MAI Cyber 2 Flash is listed as $0.10 input / $0.01 cached input / $0.50 output per 1M tokens which places it more on par with Luna/Haiku than DS 4.1 Flash. Didn't see the price for Code though.
That is very cheap, but DeepSeek from 3rd parties goes for like $0.036 in. Source: https://openrouter.ai/deepseek/deepseek-v4.1-flash-1017
>DeepSeek from 3rd parties goes for like $0.036 in
Lmao no it doesn't. That is the cheapest price and from a single random provider.
Sure, go ahead and use them. Maybe they're serving you a decent quant. Maybe they're even serving you the real model.
The absolute cheapest provider I'd ever touch is DeepInfra (they have plenty of problems) and they are $0.09 for input.
Yeah that is right, that was the lowest not average! Even at that $0.10 is very reasonable! Tbh the quant wouldn't be an issue if openrouter didn't default you to spread across different providers, makes it hard to get consistent results for evals.
These cheaper ones are usually limited on context window or token speed or latency or heavily quantised.
Click on the provider, scroll down and you'll see effective prices. Both DeepInfra and DeepSeek are at $0.008/M input. DS Flash is unbelievably cache efficient.
I wonder how much it helps that they've built their own seemingly very good AI chips, Maia 200 now Maia 300. Their blogging on these are fairly good from a spec sense (but also pretty opaque about the architectures). I don't know whether thesehttps://azure.microsoft.com/en-us/blog/deep-dive-into-azure-...
Could be mistaken but I don't think Microsoft gives direct access to customers, unlike Google tpu or aws Trainium.
Can you give an opinion on its communication style?
It's succinct and easy to understand. Some examples:
"The UI currently bypasses universal search when a ranked row is clicked, and the backend’s paginated family response can replace partial results with an error. I’m narrowing the fix to route ranked rows through the shared search and preserve the usable family when the request scope changes."
and
"Yes. JevK5 matches Qwen because Stark classifies it as the Qwen3.5 / 3.8 family.
That family metadata is included in installed-model searching at StarkApp.swift, but isn’t always visible for a single-model group. Hovering the row should reveal “Qwen3.5 / 3.8” in its details.
Its public Hugging Face metadata also declares qwen3_5_text and base_model:Qwen/Qwen3.5-4B, so a catalog match is valid too. It’s not a bug; the confusing part is that the matching metadata isn’t displayed prominently."
Very cool, thanks for the examples!
If you had to compare it to sol or opus, is it more easy to back-and-forth with? Imo opus devolves into word salad quite easily.
I thought they were in bed with OpenAI. Did that change? I'm so not up-to-date...
The OAI deal was a transitional alliance of convenience to quickly accelerate. Investing / loaning ~$10B into Open AI (much of it in compute credits) was MSFT playing catch up when they realized they were behind in the LLM race. They weren't so much worried about OAI / Ant, this was more about Google who had DeepMind and Google Brain plus significant advantage in custom TPU compute at a time when MSFT still lacked an internal frontier foundation model team.
MSFT got an IP license and a frontier model to white label along with an anchor tenant as they scaled Azure's AI compute. OAI got investment at a high valuation (good optics) and scale compute. In the meantime, MSFT quickly got to work recruiting their own frontier foundation model team in MAI.
No, for a few years now any semblance of exclusivity has been ended. They have used other model providers for a while now, including using Claude Cowork for Copilot Cowork.
That is not completely true. They have still access to IP until 2032.
https://blogs.microsoft.com/blog/2026/04/27/the-next-phase-o...
They have access to IP afaik
I think Microsoft should provide some of its models as open source/weight if they hope to secure any market share.
Otherwise, it's Windows Phone OS all over again.
These are 90-95% cheaper than Anthropic. I'd prefer open source, too, but this has nothing to do with Windows Phone OS (which failed because of network effects).
In this case there is huge demand from companies who consume API tokens at currently high prices who do not want to buy from the labs and who find that existing open source models are starting to lag the frontier too much, or are not actually that cost effective once accounting for the amount of tokens they consume.
The audience here is enterprise looking to save money, and there is no natural monopoly preventing this from gaining traction.
> which failed because of network effects
Not to nitpick, but Windows Phone also failed because Microsoft released the first generation of devices with Windows Phone 7, convinced a bunch of people to buy them, and then released a second generation of devices with updated Windows Phone 8, a version that the first generation couldn't upgrade to meaning that anyone who got in on the ground floor got locked out immediately.
When you go out of your way* to screw the people most interested in your new product it doesn't really engender large amounts of trust in what would have been your second wave of customers.
* Well, to be specific, WP8 was based off of the NT kernel and WP7 was based off of the WinCE kernel so it was probably 'the right choice' on some level, but making the entire first generation of a new product line EOL immediately rather than making an effort to ensure those people get... something... is still a dick move.
yeah they should have had a trade in program, that would have bought them serious brownie points all around.
But Windows Phone history is a long list of things they should have done
WP7 -> WP8 was rough (as was WM6.5 -> WP7), but WP8 -> WM10 was worse... while you could upgrade WP8 devices to WM10, IMHO, it wasn't worth it. Better to support old and new until the old ones age out. IIRC, WP8 had a big spec jump (single -> dual core, ram requirements, etc), so an upgrade probably wouldn't have worked...
The final build of WM10 was actually not awful... too bad it came 2 years after launch. :P
People talking about network effects killing windows phone probably don't know about how well WP8 sold low end phones in developing markets. Of course, WM10 abandoned that part of the market ... just another thing Microsoft could have done differently. :P
100% this. Lots of people want cheaper. The demand for same performance but cheaper is off the charts.
Agree. I was talking to a coworker the other day and remarked that I would be pretty content if we had five years of Claude stagnating in quality/capability but reducing in price/environmental/electrical impact.
It reminds me of when Apple switched to their own silicon in the iPhones and we had year after year of hardware with 1.5-2.5x performance compared to the previous year. We very rapidly got to the point where the processor in my phone was more power than I would ever need (unless apps got more and more bloated and inefficient, which they obviously did).
The performance is really fuzzy too, but the price you do know for certain.
Price is quite fuzzy too though when it's pricing tokens, because a model that's more token efficient but more expensive per token can be cheaper for a given task. And token efficiency is as fuzzy as performance so you're back where you started.
> a model that's more token efficient but more expensive per token can be cheaper for a given task.
True, but I'm skeptical that this is now overly used to ignore cost per token. Your comment inspired me to do a quick investigation. Looks like in my data, 92% of cost per task can be predicted by cost per token https://micjm.com/blog/cost-per-task-vs-token/
Definitely. Cost per task is nice in theory, pretty absurd in practice.
Well, it depends on the data source. Cost per task for benchmarks is absurd. Cost per task for real world usage, where the model identity is randomized (Like Arena or Agent Duel) seems better, though the tasks still are not going to be representative if you're not doing the same stuff as the battle participants.
Or even 90% of the performance but much cheaper. Many business tasks to not require frontier intelligence and were solvable two years ago.
Yeah if you're talking 85-90% of the performance at 10-15% the cost, you're going to win most of the business use cases over. That has been the broad theoretical appeal of the Chinese models, that they might provide that kind of value. Most businesses could not care less about open weights vs not, that does not enter into their discussions very heavily. They care about cost to performance. Fable 5.2 is amazing, and it's as good at eating dollars as it is at producing code.
> These are 90-95% cheaper than Anthropic. I'd prefer open source, too, but this has nothing to do with Windows Phone OS (which failed because of network effects).
FWIW it is worth noting that the GP's suggestion is more in line with Microsoft's own stated belief in the future of "unmetered AI".
At the moment, Microsoft is arguably meeting that claim with their support for and encouragement of Nvidia Nemotron Bolt, which is an actually pretty good, quite interesting, unusually-trained-and-sized-to-fit MoE model that was designed to go with the launch of the Nvidia/Microsoft RTX Spark machines.
(Good Sam Witteveen video about it: https://www.youtube.com/watch?v=WJpeW9lCSUM )
But it would be good to see Microsoft themselves distilling an open weights model.
Like any vendor with a profit motive they do have to decide the obvious, though: how do you make money from it, when right now it could cost several hundred million dollars to train (even if you adjust the training of a larger proprietary model so you can distill the smaller one from it)?
There is, essentially, no way to keep an open weights model from being copied/distilled/extracted, and IMO what that tends to suggest is that open weights models will essentially gravitate to sort of uncontroversial, "standard library" tasks.
So I think it also means that Microsoft's approach on open weights will ultimately be the same as Apple's: they will focus their efforts on LLMs that meaningfully add function to their OS and can be decapitated to make similarly useful decision models.
Put another way: given the geopolitical angle of the last few weeks, I think it is possible that the open weights coding models we have now may be among the best we will ever get, until the nearly inevitable GPU/tensor hardware glut arrives.
In the meantime, something else is clearly happening: much broader competition in the "good enough" models category, which is what we need to drive down the cost of inference hardware such that open source and open weights are feasible over the longer term.
Long term I think the economics point to having a "Linux of AI". I.e. Operating systems were the same way, expensive to make but everyone needed one and did not want to pay for one. So maybe many of the big corporations will eventually collaborate on one ecosystem of open weight models where they share the costs of training it. Then they can compete on other things like serving it or whatever.
Yeah, I think this too. It is not going to be cutting edge but it is going to be something that can be supported, reasoned about, customised etc.
I think we are not far off this with Pi, Muse Glimmer, Qwen, Gemma etc., but sooner or later we will all have to work out how to put money behind it.
I mean, nobody is really out there broadcasting their 501c that is collecting funds to build a world class open weight model... the question is does this 501c exist, and if not, why not make one? The first major hurdle i'm guessing, is that it would take actual liquidity and not funny-money.
I agree that is what it would take.
The main problem is that the open source movement has trouble even raising the money to properly pay linchpin individuals like Harlan Stenn (NTP).
Even Omarchy, which is uncomfortably effective at raising money for what amounts to a barely differentiated Linux distribution, is not raising LLM-training money; it’s barely even raising the kind of money it would take to post-train Qwen, I think.
Nobody is actually paying the tag price for Claude. Their subscription covers at least 5-20x the token cost (depends on the usage and caching).
Yes, they can change it at any time and in fact they have done so in the past. But I just want to point out that the Microsoft tag price is not competitive as it looks.
Go check their IPO filing. The large majority of revenue is from people paying the tag price.
> The audience here is enterprise looking to save money, and there is no natural monopoly preventing this from gaining traction.
Fun fact - Enterprises are just big groups of people.
If none of your people have ever used a tool because it's expensive, locked away in a some walled garden, and obscure then none of the enterprises will have demand for it. Management might force them to use some cheap model, but users will all be pushing for the models with mindshare, and Microsoft will lose marketshare as a consequence.
Apple knew this when they put Apple IIe's in schools. It's still paying dividends. Today, Anthropic and OpenAI know it and are actively hemorrhaging money trying to become synonymous with AI. Microsoft seems to think that this is all a race to the bottom. I think they're wrong.
FWIW, the frontier labs have like zero barrier to entry, even for general population. In contrast, open weight models require technical expertise to run (and/or paying third parties that host them, which requires researching and tracking which ones work and are trustworthy - i.e. hassle). So you're right in theory, but that argument evaluates strongly in favor of labs & current MS approach, and strongly against open weight.
Plenty of Open Weight models available from API providers.
There is a lot of choice...
ho do not want to buy from the labs and who find that existing open source models are starting to lag the frontier too much, or are not actually that cost effective once accounting for the amount of tokens they consume.
one is not true, the other is not all that relevant considering disparity in pricing local vs API even subsidized.
Even if both were true, this is Microsoft. There's negative chance of them producing anything close to frontier anytime soon or at all. They've been absent altogether and relied on others; The little they did is not even in also ran category like grok is for example.
The point is you no longer need frontier models to get “good enough” AI, and MS is offering a marginal improvement over open-source models at very good value
I know at least one company where people are forced to use coding agents via microsoft contract, and their monthly quotas for developers are limited to an equivalent of at best couple days of Opus 5.5.
they will happily use a microsoft model that is slightly less capable at fraction of the cost.
that'd make sense
But looking at Github's recent track record, and the outage last week, would you really depend on this offering?
Cheaper per token when used via the API.
How do they compare with the fuzzily-priced only-our-agent cheap tokens you get on a Claude Pro or Max plan?
If it were truly as good as Opus 5.5 and 95% cheaper with API pricing, then it would match what you get from a Claude plan, and I'd be willing to switch. Right now I go through some contortions to make sure I use up all of my tokens every week, and it would be nice to just pay for what I use, for those down weeks.
I'm seeing MS code quality in their products. 90% cheaper isn't worth it
Sharing open weights models is a way to break into a market where you need recognition and brand awareness.
Microsoft doesn’t need any of that. They have the market share, distribution, brand awareness, and a lot of their potential customers already use Microsoft products somewhere.
The phone comparison doesn’t apply, IMO. Phones are entire ecosystems and they’re very sticky. People don’t switch phone OSes casually because they only buy a phone every several years. Getting locked into a phone you don’t like is too risky.
AI plans are not like that. Not yet, anyway. Vendors push for contracts and one year lock ins via discounts, but individuals and nimble companies are changing AI providers all the time.
Even my wife’s big company technically has a contract with one AI provider for the company, but teams across the company are getting extra spending to use on Claude or other models of their choice and the favorites are changing every few months.
Trying an AI model is low risk. You can do it for a month or even just earmark a token budget and try it. Nothing like buying a phone that you’re stuck with for years.
Sharing open weights models is a way to break into a market where you need recognition and brand awareness.
Microsoft doesn’t need any of that. They have the market share, distribution, brand awareness, and a lot of their potential customers already use Microsoft products somewhere.
This actually sounds like exactly why they should share open weight models. Microsoft is notoriously horrible at branding. Ask five different people what "copilot" is and you'll get five completely different answers. What's the latest Xbox? I regularly game on PC and I couldn't tell you, but I can tell you the latest PlayStation is PS5. I regularly run AI models locally, but, having not clicked on TFA, I couldn't tell you what models Microsoft has released since Phi.
Microsoft is essentially nobody when it comes to AI branding. People only know of copilot, and that brand is clear as mud now.
They just need to make VS Code’s tools as attractive as Codex/Claude. Their harness layer has fallen behind. (Could they just vibe one up? Hmm.)
Microsoft already has a large market share in model serving and their AI platform is well integrated into their office-business-cloud offering portfolio.
MS AI inference is heavily used in business, especially among companies who value or require 100% assurance that their data is NOT leaking into the open. Microsoft has a secure AI interference and training platform and provides such guarantees, even for classified information.
Also: this is important -- https://microsoft.ai/news/introducing-mai-licensed-weights/
Windows phone was hardware and was decades ago... it has little bearing on their AI business model which really more of Cloud offering. Microsoft is doing very well in Cloud services and revenue... Depending how you slice up their cloud services business, Microsoft is the largest (by a fair margin) or second largest cloud computing business in terms of annual revenue.
All in all, if the models perform as well in the real world as they do in benchmarks, this is fairly impressive and also concerning for OpenAI and Anthropic (downward price pressure).
If you don't want your data leaking into the public, then you should probably not let Microsoft stuard any of your data. Azure's security model is based on imagination (vibes in modern parlance?) and should not be relied on for serious work.
Knowing Microsoft, they would rather let their closed source OS die than open source it. That's rooted deep into their core DNA and that's exactly what happened to their phones. You can't really change who you are, fundamentally.
Their phones failed because other companies like Google and Apple banned their apps from being in their app store - effectively making it impossible for either Android or iphone users to switch. MS also made it really hard for developers to get their apps in their app store as well. Submissions form failures, long approval processes and bottlenecks with rigid certification rules all contributed to the Windows Phone failure.
It had nothing to do with open sourcing their OS.
Its important to note most of the groundbreaking features the Windows Phone OS had were imitated by both iOS and Android in subsequent years after its failure.
You mean the "groundbreaking" features from Maemo/MeeGo in Nokia N9 and even some from Palm Pre WebOS?
You would be surprised how many companies only allow their employees to use Copilot simply because they have existing agreements with Microsoft and it takes too much to start a relationship with one of the labs.
The big advantage I see is you can run Claude in Copilot- while also getting full access to Sharepoint/Outlook/Teams. I can reference a conversation in email and get instant context.
True in my case. My company and the main client company I do work for give developers Github Copilot. You can't just install Claude. I see these MAI models available at 50% off.
Anything short of the absolute frontier models are now essentially fungible.
Price/token and data protection agreements are all that matters.
MS dominates at the latter, and they can subsidize the former. I would not bet against them.
Some of their existing offerings provide finetuning service and I guess that'd be sufficient for enterprise customers who're locked into the Microsoft ecosystem anyway.
This topic has its own dedicated section in the linked page.
It's not enough for companies outside the US, though.
They're not open weights but they're open... something? Kind of sort of?
> We’ve heard again and again from leaders that no one wants to pour their IQ into someone else’s model.
> We heard you. We started with Custom Models. AT&T developed an MAI custom model for 5G network operation analytics and it outperformed GPT-6.1 Sol at 50x cheaper cost. On complex scenarios, they found that Code 2 delivered frontier-level performance at a significantly lower price than alternatives.
And a link to: https://microsoft.ai/news/introducing-mai-licensed-weights/
I don't really understand what this means.
Means if you have deep pockets you can license the weights and do your own training with them.
I don't think you can compare to Windows Phone. That was a vastly different situation dependent on hardware partners, cellphone networks and network effects in terms of apps, etc. etc.
If anything this highlights how commoditizable LLMS are. Is it as good as Opus but significantly cheaper? Then that's a very easy switch. Not to mention Office and Windows... they're not short of ways to get this in front of people.
Commoditize your complement.
But you have to choose which is your complement and which is your main product.
It seems to me that secure inference and horizontal integration is their main product, and the complement is the LLM itself.
> if they hope to secure any market share
Google and Microsoft are already the bulk of the market share.
Most knowledge workers are either on Microsoft 365 or Google cloud services for anything from mail to meetings, calendars, calcs, etc.
And all of Gemini and Copilot have already moved into the offices of most people.
Why is hackernews absolutely obsessed with open source as a pathway to success? So extremely few open source projects have been successful, while almost all successful products are not open source. None of the most successful current AI models are open weight, so why should this be considered a factor in success?
How do you define "success" and "extremely few"?
Linux, Android, ElasticSearch, Apache, Nginx, PHP, Python, Perl, WebKit, Mozilla/Firefox, Chromium, Postgresql, Rust, MySQL, Redis, React, Jquery, VueJS, Blender, RepRap/Prusa... I am sure I could come up with a 100+ off the top of my head, I was just mixing it up a bit.
Because we want open source for all the usual reasons: freedom, price, availability, community, etc...
Open source makes the world better, "successful products" make a couple assholes rich.
This. I literally do not care about any of their offerings unless they are open weights.
I'll pay Microsoft to run them if they're open weights. I don't want to deal with the hassle of doing that myself. But I won't even touch them if I can't possibly own the weights myself.
The market is already full of closed weights. We don't need more.
You have stated your position but not explained it. How does it help you if the weights are open?
Risk.
Can i have more explanation please? I'm not in risk management but i would like to understand the situation better. Just risk that they change it under you? Or that they're training on your data, or what kinds of risks are we most worried about?
They can make your model disappear. If you have a system and a set of prompts and evals and everything is working well, it's no fun for that model to just disappear. Some will reply "do a better job building your system so it's not model dependent" and they are sort of right, but that isn't always economically feasible and/or just isn't the reality of how a lot of systems get built.
Models hosted as a service can be subtly tweaked without any announcement, or suddenly be 'upgraded' to a new one with a ticking clock on how long the old one sticks around. In both cases, the flows people establish around them are disrupted or broken.
I've lost track of the anguished cries of those unhappy when either of these has happened.
That's a problem with decision models, for example, but coding models have no such risk. I switch between multiple coding models from multiple providers each day.
You must be new here
History shows that if you give a corporation power to extract money from you, most of them will do that without any regard to the user.
That power is the model weights. If there is a market around open weights, they will exist and corps will not be able to squeeze so hard.
Unfortunately not many people understand this simple and intuitive reality, and thats why we suffer
This assumes material switching costs. I use frontier max plans for now but have fail over models for each class of model action with tuned context so I can keep the performance when I lose the max plan subsidy at a lower price that staying with the frontier labs for certain classes of work.
I see minimal downside in using closed weight models as long as they're behind a router, you have good prod sourced evals and you have fail over to other models configured and tested (running either dark or active:active) with real prod workloads.
If you want relatively repeatable and predictable responses (even within the reality of being non-deterministic) then making sure you'll have continued access to a model to replay responses for debugging and tuning can be important.
Yes, we do need more. I would like significantly lower price for same performance. I'd like higher throughput.
The Ignite keynote looked so shallow and opportunistic to leadership. You wonder if they have any clout anymore, and it feels like the engineers have abandoned the ship. I am working day to day with AI Engineering, and it feels I have zero reason to do business using any of their tech stack. This clearly started with the dev division re-org, they completely missed the memo. There is zero innovation, and it is more than time for Nadella to go.
Microsoft marketing and branding dept is so damn lazy. They just name things from stuff in their environment, the educational channel was "Channel 9", the same as the local PBS station, KCTS. And the Ignite conferences were started in Seattle, https://en.wikipedia.org/wiki/Ignite_%28event%29 all MS does is uncreatively steal. It is embarrassing.
> and it is more than time for Nadella to go.
I'm glad I'm not alone in thinking this. I thought the same about Tim Cook, and some people were angry at me about it, but sometimes you need to pass the torch and let someone else have a go at it to get a fresh vision, Apple could have been further along with AI in my eyes as a solution for running it locally, but they are so risk adverse since the days of Steve Jobs, I want to see Apple be slightly bolder, not just "marketing" boldness.
I think Microsoft needs some serious surgery top to bottom, and to bring back top engineers, there's no reason Google or Microsoft should have taken this long to be comparable to Anthropic and OpenAI, they have historically had highly talented and capable individuals so it has to be the leadership.
If AI is truly their focus, can they stop cannibalizing Windows and make it a less awful OS with less cruft on it? The amount of memory usage on Windows vs Linux or even macOS always astounds me.
Counter point: Apple staying out of AI and simply piling cash while their competitors burn it is the reason their stock price is soaring and may turn out to be a very savvy business decision.
And about running local models, they seem to be the number one recommended option for that? I don’t have much knowledge around local models though, I may be missing something, but I would guess their hardware sales have increased because of AI.
Theyre just the number one right now because they happen to have shared memory architecture. Which is being done now by others.
Their long standing feud or whatever you want to call it with Nvidia is kind of screwing them however
Agreed - Apple may not have a frontier model, but they're positioning themselves to be an AI powerhouse. They'll eventually own the chips and the software - unlike anyone else in this space.
In typical Apple fashion, they're not rushing into a bubble, but playing the long game.
To be clear, what I would have loved to see more from Apple is local AI inference being farther along, and to be extremely fair to them, what they have achieved is amazing.
It’s second Ignite keynote in a row that is totally unimpressive and uninspiring. At least Satya Nadella came back to show he still runs the vision of the company (whatever the vision is). last year his COO entitled as some minor CEO of business ran the whole show and it was very confusing.
They still suck at AI vendor and struggle to evangelize that they are just an AI integrator with something like AI-Biztalk that is not easy to position as legitimate product.
It was interesting to pitch CoPilot as AI-OS idea… but they don’t look like they can succeed converting it from buzzword to reality and swarming agents is something they can hardly monopolize…
I do believe they are preparing a succession plan for Nadella thus the second CEO not a simple COO like with Ballmer-Turner duo in the past
Unless you're already in their ecosystem, why would someone join microsoft? They are becoming like Oracle. Their end user devices are avoided at most modern start ups, their cloud is expensive and confusing, their cloud customer support is awful as well.
Microsoft is also:
- A big contributor to Postgres OSS work
- A big contributor to Python
- Ships VS Code and Monaco, on top of which many, many startups have built their products and VS Code (or offshoot) is used by every startup I've been in and increasingly seeing adoption in enterprise
- Ships TypeScript, a cornerstone of the modern JS stack
- Owner/operator of GitHub which, despite its issues that we all know of (no point trashing them here), still holds a massive network effect with the majority of open source projects hosted on GH, running billions of Actions minutes/mo. I'd guess.
Microsoft and Oracle are very, very different.
> VS Code (or offshoot) is ... increasingly seeing adoption in enterprise
I think a solid 90%+ of developers I know at my f500 company use VSCode.
I started noticing another trend - people move off VSCode because it became too bloated and slow to start. Built-in copilot is one of the reasons. Current models are quite capable, so often you don't even need to open the IDE.
Yeah, I have it disabled and only use the terminal for LLMs for that exact reason.
Oracle is the biggest contributor to Java, and all new releases are superb.
Also top contirbutor in the linux kernel and does most work on MySQL, which is... in much better shape than it was under sun.
Well said!
For one their online productivity suite for companies is cheaper than google and comes with actual support if something goes wrong.
You just sign up, yeah you have to put up with their login and UI once but then you have an API key and you are done.
Why are they not listed in the Artificial Analysis index?
Maybe they didn't make their model available for benchmarking.
This seems a pretty bold statement, especially the Opus 5.5 comparison.
"MAI Code 2 reaches 98% of the coding quality of GPT-6.1 Sol at about 40% of the price. And against Opus 5.5, it reaches 96% performance for just 5% of the price."
> 98% of the coding quality
Can someone help me understand how they arrive at this figure?
98 can be bad. E.g. uptime or aviation safety.
98 can be good I.e. human test score or funds recovered from debt.
I think for AI tasks it is somewhere between but 2% losses will add up (or you need more loops to correct and hence cost)
Dont worry - it only hallucinates security breaking bugs in 2% of the code it produces. Comparatively :-D
In all seriousness I have no idea how they came to that conclusion. We really need better standards here for evaluating these things but obviously that is difficult.
I think they mean "98% of the average score on these old benchmarks which are no longer representative".
5% of the price compared to Opus 5.5 Max reasoning, and that's at a 50% promotional discount for MAI Code 2, I believe.
With Opus 5.5, Max would be around 4.5x more expensive than Medium reasoning.
That MAI Code 2 is within 4% of the performance at Max seems unlikely, to say the least.
Actual performance and a price comparison to Opus 5.5 at Medium (Anthropic's recommended setting) would be interesting.
that's the thing with coding quality. as we all know, one line of code can ruin everything, so every small increment matters much more than the cost.
How much % of price to reach 200% of performance?
>98% of the coding quality of GPT-6.1
Meanwhile, Deepseek 4.1 flash is 3000% better at 3.5% of the cost, in that it gets work done (I have never had anything really done with GPT 6.1 Sol).
This is real token usage in a complicated monorepo codebase for adding a feature:
---
315,207,850 tok
Cache hit
99.6%
Uncached input
1,187,911 tok
Cached input
313,640,960 tok
Output
378,979 tok
---
It ended up creating a few PRs, with a total of around 6000 lines of code and the feature works as expected. Total cost was around $1.35. GPT 6.1 Sol would have cost around $40 for this and from my experience with it, it would most definitely have had so much bugs that it would not be worth fixing.
So, Microsoft would charge you $16 to waste your time even more. What are we even doing?
Microsoft and OpenAI are both targeting Anthropic where it hurts most: high token burn and cost. Microsoft is a bigger threat because they already have the hearts and minds of enterprise for decades.
Plus they have their own cloud and compute and they have access to OpenAI.
I wonder if the cost is artificially lowered (using lower cost they set themselves for something) or if there are an objective calculation there.
Other providers could also serve at these costs. It’s truly a race to the bottom to win market share and all of us get to benefit. A rare treat.
> and all of us get to benefit
Let’s hope it isn’t a temporary benefit. If the cost to train models has improved to make the cost of use more affordable then it’s certainly a win. If instead the cost of use is merely being subsidized to undercut competition with plans to increase later after majority share is won, we may find ourselves in a tough spot.
My guess is that training of new models will slow down massively post IPO. Claude/Anthropic are racing to be the most impressive until then, at which point they need to start generating some profit. The models are getting to the point where they're so powerful already, of course they can always be better, but honestly most consumers would prefer more speed and lower costs rather than better models at this point.
Yeah these models aren't necessarily any more efficient than OpenAI or Anthropic's or Googles models as noone releases weights/model sizes.
They haven't and likely won't release weights but their research blog posts suggest they specifically set out to train models for efficiency: https://microsoft.ai/news/frontier-pre-training-at-microsoft...
That was my thought too. They are probably eating a ton of the operational costs just to gain market share. Then they too will probably have to eventually raise their prices. But have no proof for this (except that no other US AI companies can claim rates as low as this).
I don't think they are. They explicitly changed GitHub Copilot billing in June to essentially match Api pricing. And their research posts for these models suggests they're specifically targeting efficiency and designing with/around their custom MAIA training/inference chips. https://microsoft.ai/news/frontier-pre-training-at-microsoft...
Models alone won't be enough. The GitHub Copilot usage-based pricing change really damaged the relationship with so many developers. They would need a much more enticing plan to make people switch from Claude or Codex. Or maybe, they just don't care about individual developers. B2B is all they need.
in enterprise pricing, also anthropic and openai are usage based. So nothing was damaged there.
Enterprises move slowly. They should be trying to gain mind share with individuals. Anthropic’s non-usage-based individual plans allow as much as ~$13k worth of tokens. What is Microsoft’s answer to that?
Lots of comments around Anthropic's moat here and how its non-existent. I bet anthropic is going to build a moat around two things:
1. brand recognition: everyone is going to simply default to using claude because thats what they know and recognize.
2. Applications for Business: Anthropic is going to go higher up the stack and build the specific apps for each business - Legal, healthcare, manufacturing, finance etc
The models will be the platform things are built on and the apps are the things that people get locked into.
I do default to Claude and it would take a lot to stop me from doing that. However, 98% the quality at 5% the price is a lot and if that's true I'll be switching to mAI.
At least in those cases where I'm paying API pricing.
The argument about the lack of a moat around AI has been put forward several times here, this thread is nothing new.
What has been evident, for some time, is that the only real moat is who is prepared to pay the massive costs of being in the race - Was it Anthropic, or OpenAI that posted a spend of 80 billion for the year.
If anyone is going to "win" it's not based on whats available now, not by some margin, instead it will be "reliability" (effectively, how good the model is, how much can we actually rely on its output) and, cost.
The cost of running these things has to drop markedly for them to be of any real value. Right now we are in the VC buying the market era, which subsidises the cost of use. Once people have to pay the real cost to use these things the "market" will drop away, very quickly
margins on inference are very high tho. at some point ( soon?) the models will be "smart enough" for bulk of tasks. so this current regime of crazy traijing costs should end
I appreciated the concluding remarks at Ignite:
> Frontier transformation is intended to enable human ambition. ... We’ve given AI models far too much credit in the intelligence category. ... Intelligence means something more. Your unique ideas. ... Let’s all stay grounded on the principle that AI must do more for humanity.
What are you appreciating about this clear marketing bs slop?
Perhaps it's insincere, but I believe in the substance of the message, and I'm glad it's at least being presented.
They've held this position consistently and for a long time. It's how MS is approaching AI in general. It's a tool, not our next overlord. At least, the ones from MS. Altman would be happy to follow Astra into the wilderness, I think.
They haven’t and they don’t though. Microsoft PR statements have always where ever they think the currents are flowing.
At the same time, Microsoft put copilot in MS Paint.
It’s about time Microsoft punched back. These are actually pretty good. Not just for a niche use case either (Phi).
From a business standpoint, Microsoft should ideally not choose to compete with OpenAI as they own 27% of OpenAI, and they might never gather enough capital in terms of both money and talent to be competitive with OpenAI or the other 2 big western AI players in the form of Anthropic and Google (as there’s a rule of 3 in capex heavy markets, and in the case of Google Search that it almost fully monopolized it).
At least for example Mistral has a case of sovereign AI for Europe, a niche which Microsoft cannot fill.
How am I supposed to use this if not via the more common inference servers such as OpenRouter or Bedrock? So many model companies are making this mistake nowadays thinking anyone who wants to try their model should have to sign up for some obscure proprietary API
Most likely only targeting large enterprise B2B clients and trying to get them to sign longterm contracts before the cost of all models goes down drastically. They may even be subsidizing these current models and hoping longterm their next gen of model will be cheaper and better as has been the case so far.
They will be releasing these soon on OpenRouter, Vercel, Hermes, OpenCode, and Kilo.
Why are these new models not on Artificial Analysis?
Hi,
John here from the team that built Microsoft Decision 1 and 1.1.
Microsoft Decision 1.1 is on the Artificial Analysis Decision Index v2.1 leaderboard, where score 68, above Gemini Decision 1, GPT-6 Luna Decisions, and Jev 1.14. [1]
We also compare ourselves to other decision models in our blog post on Microsoft's Command Line blog from November 17 [2]. We name all of the public benchmarks we evaluated against in the post, and provide a breakdown by category.
[1] https://artificialanalysis.ai/evaluations/artificial-analysi... [2] https://commandline.microsoft.com/microsoft-decision-model-g...
this doesn't bring a lot of confidence when you are using old version of Terminal Bench or Frontier Code when new variants were available for quite quite some time.
I could be wrong, but I think Artificial Analysis typically runs benchmarks independently after models become publicly available.
A lot of times they are given early access so they can release benchmarks on the day of the model public release.
But the odd thing is that even the previous generations of these models aren't on Artificial Analysis. It's been months since version 1 of some of these models was announced, but none of these are on AA. Are they only providing them with anti-benchmarking language or something?
I can't find any reports of people having used and reviewed these models; it looks like they didn't release it to any reviewers before this public announcement. That's not a good sign for whether the real-world performance will live up to the benchmarks.
And, honestly, Microsoft is a severely toxic brand. I'm not a fan of Anthropic or OpenAI, but at least they pay lip service to safety and don't have history of consistently bad behavior going back decades.
I think if they announced MAI Code 2 Flash to be not just for coding , but for other applications as well, it would do a lot better considering its 65% the cost of GPT-6 Luna.
Yes, for general language tasks, Google's models win. I wish we had more choices in non-coding space, but even the locally hosted Gemma4 models win against the cheap Sonnet / Luna options.
Still just waiting for one of these smaller-but-good-enough models to be bundled in with the OS so I can build desktop apps on them that don't have weird charge-the-developer-the-more-users-use-it financials...
Are there any numbers for Maia? A recruiter reached out to me for their compiler team and the hiring manager wasn't able to name anyone on the team he thought I might know based on my resume. Best he was able to do was a guy I worked with 20 years ago left their team a year ago.
It has all the markings of a classic MS project where the implicit assumption is a half-assed attack is enough, they just need to tie it to something popular.
I found their transcription model to be one of the best i have used
Dumb question, but how are you accessing them? OpenRouter? Azure?
For a brief moment I was really excited that we would have a cyber-capable model at this price range coming from a legit provider, but I was disappointed to see that they won't serve that product outside of their beta product.
Which beta product?
so how do I use these exactly?
That's what I wondered too. I know how to sign up and use models from other companies, but it's not obvious at all where to do the same for MAI products.
They're available in the model list in Github if you have a Copilot subscription. If you're part of an org you can turn them on here and use them in VS Code Copiliot extension: https://github.com/organizations/<your-org>/settings/copilot/models
This tends to be typical of Microsoft as a whole. It's like they put effort into making everything as unintuitive as possible - UIs in particular. Then once you've finally learned to comfortably make your way around a platform they move everything again.
They will soon be available via OpenRouter, Vercel, Hermes, OpenCode, and Kilo
It's in the Github Copilot app, and extension in VS Code/VS right now, and will be added to openrouter soon.
Happy that this time they are releasing them beyond GHCP too. Looking forward to test the flash one.
Looks like they aren't on Open Router (yet). Very excited to give them a spin on my evals once they get over there.
I don’t think it will take much time for them to reach openrouter, it’s definitely in their roadmap.
What harness can I use which will allow this model but also be able to switch back to Anthropic or open models?
GitHub Copilot! https://github.com/features/copilot
(well, at least the MAI Code models for now. They recently rolled out decision models and image models though, so I imagine the others will follow!)
Any byok client I assume. I like agora on android and aichat on linux
If they offer it along side with opus/sol and show price/token used side by side, they could win some share. But they have to be confident it performs, otherwise people will look at it as Sol vs Luna or Opus vs Haku if it sucked.
MAI Image 2.6 Flash was about 2c / image, and it looks like 2.7 Flash is 1c / image. I used 2.6-Flash as a backup when my low cost Krea2 provider flaked out and the results were surprisingly ok (comparable with Krea2 Medium Turbo, cheapest provider cost is 0.67c / image).
At 1c it might even be a default choice, I'll definitely try it out.
> MAI Code 2 reaches 98% of the coding quality of GPT-6.1 Sol at about 40% of the price.
To me, this smells like the same trick I always see in claims about model routers. As in, running against a saturated benchmark then showing huge cost savings for a tiny performance drop (because almost any model will perform well).
Looking deeper, though, it looks like they are citing four benchmarks, only one of which is basically saturated (Terminal Bench 2.1).
On the other hand, on 2/4 of these benchmarks Sol 6.1 does better than Opus 5.5, which just doesn't pass the sniff test, so my conclusion is that these are cherrypicked benchmarks.
"Decision 1.1 is the latest decision model from Microsoft, 50x faster and 79x cheaper than GPT-6.1 Sol"... uhm okay. What a useless comparison.
They're talking to executives not developers.
I had this thought yesterday, the era of developers picking tech is probably over. Microsoft's wheelhouse is selling to upstairs.
They're going to eat everybody's lunch that is still playing the old game of selling to the devs.
> the era of developers picking tech is probably over
Don't really know when that era ever existed
It still exists, just not in time, but rather at the company size.
Has this been true pre-AI? It has always been sold to executives, no?
It was a short era
I'd be curious to hear how AI purchasing is really significantly different from other tech purchasing decisions. Seems like the enterprises that are accustomed to strong-handed mandates from on-high will continue to make tech decisions from the corner office, and the smaller orgs or more heterodox ones will continue to listen to ICs and management who surface new tools.
Even if they are talking to execs this comparison is still maliciously misleading. If facts and truth don't matter anymore anything in this age then why not add a few more zeroes to this number...
It can't fail to meet expectations immediately, only after it's been sold to execs, they present to the board saying how technically savvy and fiscally responsible (compared to the dang engineers asking for claude), and then once it's worked into the ecosystem it can slowly become more well known how bad it is. The art of the deal as it were.
> Microsoft's wheelhouse is selling to upstairs.
It's the modern era version of "Nobody ever got fired for buying IBM"
Hi,
John here from the team that built Microsoft Decision 1 and 1.1.
We compare ourselves to other decision models in our blog post on Microsoft's Command Line blog from November 17 [1]. We name all of the public benchmarks we evaluated against in the post, and provide a breakdown by category.
We also are on the Artificial Analysis leaderboard, which is an independent third party assessment. On AA, we score 68, above Gemini Decision 1, GPT-6 Luna Decisions, and Jev 1.14. [2]
Our previous MD1 model from October is also benchmarked on JevBench, another independent benchmark for decision models. [3] On JevBench, we are 9th place for API models, ahead of Gemini Decision 1 and Jev 1.14. We look forward to seeing how the new MD1.1 scores on JevBench.
With regards to cost: Our models are the same price as Jev on a per-token basis. However, on both JevBench and Artificial Analysis benchmarks, our models use fewer tokens per task than Jev. For example, our cost score for MD1.1 on Artificial Analysis at $140 per 1M tasks compared to Jev's $156 and Gemini's $345. There are similar stats for latency, where our new MD1.1 model is 2x faster than our previous one.
[1] https://commandline.microsoft.com/microsoft-decision-model-g... [2] https://artificialanalysis.ai/evaluations/artificial-analysi... [3] https://benchmarkheaven.com/jev-models/api
A little bit further down the page:
> Decision 1.1 provides model routing integrated right into your applications so that the bigger, pricier models are only used when you really need them. On the AA Decision Index Microsoft is ahead of Jev, Google and OpenAI, while being faster and cheaper. Used by 4,000 customers, it’s 50x faster and 79x cheaper than GPT-6.1 Sol.
Tangential to this, but I really hate the use of “#x times cheaper” when describing a price difference, especially when it takes place after an “#x times faster” statement. It’s initially confusing and makes it harder to mentally calculate the actual difference. Just give it as a percentage!
Agreed, and it's also incorrect - "1 times cheaper" would be free. "79 times cheaper" is nonsense.
It's the inverse of 79x more expensive, no?
I think that's what they probably meant, but it's not exactly right.
Consider something being 20% cheaper - it's $8 instead of $10. 90% cheaper means it's $1. 100% cheaper is free.
Also imagine two products - A is $10, B is $20. B is 100% more expensive than A, but A is 50% cheaper than B.
(79x is 7900%, of course.)
What do you mean "the actual difference"? 79 times cheaper means costing 1/79 of compared price. That's what it is. Or do you mean you want to see it in the format of decimal fraction?
Instead of “79x cheaper” say “1.26% of the cost” so readers can quickly do a mental cost comparison
Which comparison is the question.
"My budget gives me 100 million tokens for GTP 6.1 SOL. How many tokens for this new model would I get? Let me quickly divide by 0.0126 in my mind to figure it out."
x times more/less is the simplest form of comparison because bigger value is divisible by lesser value. I'm saying this because we first teach children division and multiplication and only then fractions.
Our banana is 9 times yellower than the competitor's oranges.
It is also substantially better on most frontier benchmarks than a bash script with a random number generator.
Yeah...they really should be comparing that to Jev.
Makes me doubt the rest of their claims when you see something like that
Why? Look at their list of models again, each is specialized to a specific task. None of them are general purpose. It's not surprising that a task specific model can be better optimized for its task than a general purpose model.
They are acting like Sol 6.1, a general purpose model, is comparable to a decisions model. OpenAI has a decisions model, so they should be comparing to that.
> Today, we’re taking a big step forward with MAI Licensed Weights: models under your control.
>Take a frontier MAI model, customize it using your proprietary data if you want to. Then run it securely in your own Azure environment.
Ohh Microsoft. Is it really my environment if it’s Azure ?
Anyway I wonder if GitHub Co pilot will start auto routing to these new models. All those gamed efficiency metrics sure do look impressive.
This whole article is strange. It’s as if Microsoft is seeing the Anthropic IPO and saying “um we have models too”
This whole week has been like that. Open AI earlier this week, MS today. Presumably Google tomorrow.
I wonder if we’ll see a wave of free tokens to try and get people to switch.
Is it really my environment if it’s Azure ?
I take your point but for lots of enterprises that already use azure this will be attractive
"from the company you can trust."
Well they're not wrong... if you trust them to have telemetry, take their own Github repos offline without notice or reason, exceed their Azure SLAs, and break Windows installs on a monthly cadence.
>This is Frontier Intelligence you can trust, from the company you can trust.
Pretty bold, after what happened last week. Race to the bottom wrt price-per-token is accelerating, and glad that large US corporations are now joining the race for frontier open source models. Look forward to open source releases from Google (Gemma has become stale now) and Meta (Llama is the skeleton in swimming pool meme).
A new challenger appears! Good to have a new horse in this race to the bottom. At those prices and performances microsoft seems poised to get a good chunk of the market.
>Security, alignment and containment are the most important priorities for agent development in 2027 and beyond.
Anyone else read John Lilly's stuff on SSI and laughed while reading that? Or at least, I imagine SSI is laughing ;)
and the crowd goes mild
I haven't been this excited about Microsoft products since they released their new Version of Editor for Windows 11 with support for multiple(!) undo steps in 2022[1]
[1] https://blogs.windows.com/windows-insider/2021/12/07/redesig...
I think that's just built into the browser's textarea element. (It's a web app right?)
It is not.
>MAI Code 2 reaches 98% of the coding quality of GPT-6.1 Sol at about 40% of the price. And against Opus 5.5, it reaches 96% performance for just 5% of the price.
>MAI Code 2 Flash matches the performance of GPT-6 Luna at 65% of the price, and Haiku 5.5 at 20% of the price.
These are quiet some claims!
> Claude, search through every benchmark and comparison model combination to find the best performance vs. cost claim we can make about MAI Code 2.
Amusing that this was announced 1 day before the Anthropic IPO
Which one of them wrote the KB breaking printers ?
Where is Simon and the pelicans? I for one miss them! Don't make me go try all these models for myself!
Haven't used any Microsoft technologies for over 10 years. Why is this company still in business?
They are the most diversified of the big players, and some of their strengths are in areas like Enterprise, areas which are not well understood by the hn crowd
Me neither. Market Cap: 3.75T. Clearly our use isn't correlated with some extremely large software markets (aka enterprise IT).
I haven't driven a Ford car in over 20 years. Why are they still in business?
Because Chinese cars have been tariffed to hell in their home country?
Has anyone validated these claims from Microsoft? Cuz if true wow
Given their access to data, code, hardware and research power, it's puzzling how Google and MS gave away the lead to some startups. So quickly.
Closed weights, don't care. China's still beating you.
Is it world class because someone says so?
I believe they assert that claim by citing benchmarks against other "world class" models. It's not a term of art it's subjective.
The artistic citing of world class is it's own artistic expression.
I just want my email and cloud to stay working.
Release the weights you cowards!
And open source windows and office, while you are at it ;o)
"world-class" is one way to say "not-frontier"
Terra is world-class, Astra is star-class.
Wait, no, Sol is star-class too. I'm confused now. (Also, is a moon a world?)
is MAI a different team than the one that does Phi?
> Decision 1.1 is the latest decision model from Microsoft, 50x faster and 79x cheaper than GPT-6.1 Sol.
No one compares decision models with Sol!
I guess MS thinks that enterprise executives do.
It will look like its useful. There will be business brochures about its being useful. Some sad engineers at sad megacorps will be forced to use it.
But at the end of the day: it will be what all other microsoft products turned up to be: an utter waste of time.
Spend your time on Claude, Gemini, Kimi, Trinity or even Deepseek. And thank me later.
No local models? Meh...
they wouldn't want us not to pay them on subscriptions, would they?
Weak models, closed weights, no technical information
Mostly PR stunt the day of Anthrophic's IPO.