I'd hate to be OpenAI and Anthropic. They are better than the rest of the available models currently, but only by a few months.
That is a very tenuous lead and doesn't speak to much of a sustainable competitive moat for either company. The have to keep sprinting as fast as possible on the hamster wheel to just stay ahead.
Claude has pushed hard on productization. Day to day I use codex for everything, I went to watch my wife use Claude and the way it works is so different. I was shocked when she was modifying an artifact through the Claude website and it was able to make changes to a local file on her desktop.
Microsoft didn't win spreadsheets by having the most advanced or releasing the most often. They made a really good, easy to use product, and got it in front of everyone. (Though Office's p-code was a huge advantage to portability)
I think Google and Microsoft have the benefit of time as you say, and already being profitable without AI.
Honestly, this line up is impressive, I wanted to try MAI since v1 came out, but Microsoft locked down Copilot subscriptions. If they gave me a comparable plan to Claude Max I'd consider it.
I would counter argue with "But chinese models use distillation of Frontiermodels..." which is different than being a Frontierlab or not clear how different it is.
But Microsoft coming from the side like this might be very different if it holds up.
That isn't countering though. Pointing out that distillation makes it "easy" for other labs to catch up, in a matter of months, supports the lack of a moat.
It's like they're runners drafting behind the lead runner. It is easier to run in their wake. If the lead wants to stay in front, they have to keep running into the wind.
For now I have not seen any credible evidence for "But chinese models use distillation of Frontiermodels...".
Of course, I would not be surprised if they also do this, mainly because after OpenAI and Anthropic have succeeded to gather most training information that was easily accessible in the past and now it has become much more difficult to do the same. For some information that was hoarded by OpenAI and Anthropic, public access may have become impossible otherwise than through their LLMs.
Nonetheless, distillation cannot be the main source of improvement for the Chinese models, among other reasons because even assuming that they may have access to a great number of fake accounts, there still is not enough time for a lot of distillation, taking into consideration the timeline of the releases, and it is not believable that the attempts to do distillation would not be detected quickly by OpenAI and Anthropic, leading to the banning of the fake accounts. While OpenAI and Anthropic provide negligible information about their models, the Chinese AI companies are much more transparent about their methods and we can see that there is a lot of diversity among their models and a lot of innovations with each new version.
Speaking of Thinking Machines, that's another data point. Their Inkling Small 2 model is just a bit behind DeepSeek V4.1 Flash 1017 in benchmark, and in my tests, that seems to be accurate, but mostly because they used less training data for pre-training, so they have less of a knowledge base, while similar reasoning capabilities. (One of the most interesting parts of Inkling Small 2 is that their training must have been very cost-efficient, apparently using less, but higher quality training data.)
And we're now seeing more and more Western near-frontier models from labs other than Anthropic and OpenAI to question whether distilling reasoning traces is really much of a difference maker.
Personally, my theory is that the gap is primarily due to Chinese labs having a harder time getting the necessary compute (especially top of the line GPUs) for training. These are, as far as I can tell, orders of magnitude differences in availability, so it wouldn't be surprising that a large compute gap is reflected in capabilities (and also indirectly visible through model sizes).
I don't think there is perfect proof, but its taken as truth.
Indication for this are for example that Chinese models think they are another model: "chinese model think its claude".
My assumption is, that Chinese model found things Frontierlabs already had but didn't share. There was this paper were i had the feeling that this was a real and significant Research Result from Deepseek "DeepSeek-V4.1-Flash: Pushing the Limits of KV Cache Compression" but otherwise its just unclear.
The main issue is, that as long as it seems plausable that they do distillation, its hard to determine how big the gap is.
Not saying they can't catch up but they are also potentially limited by resources (money & nvidia GPUs)
> Indication for this are for example that Chinese models think they are another model: "chinese model think its claude".
There is evidence that some Chinese labs have been training on frontier model CoT (in particular from the analysis of reasoning traces [1]), but what models claim about their identity is pretty much worthless as evidence. LLMs are not self-aware. These claims come from whatever is in their training data, and Claude/GPT output including such statements is all over the public web, including in some huggingface data dumps. Consider that people got Opus claiming to be Qwen or DeepSeek when asked in Chinese, and you'll see how weak this particular evidence is.
So you’ve been completely ignoring all the work on introspective awareness, situational awareness, training set ordering awareness (https://arxiv.org/abs/2509.14223 & https://arxiv.org/abs/2403.09613), subliminal transfer (https://arxiv.org/abs/2606.00995), workspace formation, learning mechanics/theory of grokking, etc? Like it’s ok to be ignorant about the state of knowledge but you should at least correct your flawed mental model. The “objective” that mid and posttraining optimize is so powerful because pretraining already induced factorized worldmodels which have so much predictive and generative power that they snowball with even mild trajectory optimization (very different training objective than pretraining).
(Functional) small-s latent self awareness is a reliable shortcut (this is a principle you can derive from proof theory, the metric is deductive distance and that induces geometric morphisms). State compression from accurate selfmodeling pays rent.
The frontier moves fast, but "LLMS are not self aware" is so foolishly far from the truth and has been for years for anyone actually keeping up with the situation. By now the basic science explaining how and the extent to which this is wrong, is already done.
LLMs lack reliable verbalized self knowledge, and constantly underestimate and undercut their own capability.
That's a different problem.
All it would take is Nvidia finally releasing the 6090's instead of trying to milk enterprise GPUs. I don't understand the blind love for Jensen, he's clearly been screwing over all the consumers who made him and Nvidia wealthy.
Someone somewhere really needs to build a RAM startup that not only makes RAM but disrupts the stronghold AI has over it, it's a little absurd.
I currently do not bother with chinese models. I try them out, i use the small ones for small experiments but the price point difference matters so little that I want 'the current best'.
And 6 month feels still very different.
Its for sure an option or might have some kind of impact but if you have some kind of restrictions on what you should use and the price is not that big but the quality and possibilites are better, I can see companies just staying with Anthropic or OpenAI.
You also don't just switch out a model, they might need prompt updates etc.
I disagree, I think distillation hides the true technical advantage since you can always get within X% of Claude by ex: training on newly created code on Github, but the underlying intelligence is harder to replicate.
based on the behavior of the chinese labs it seems like training on chain-of-thought is extremely important for distillation, and you can't get that by training on llm generated text.
Openai and Anthropic are better at making models easily available too. Both Google and Microsoft are using promotional pricing to push their models and that should tell you how far they are from actually competing.
Also model availability is not easy. Google newest models are unavailable on some pro accounts and perform best only with Antigravity. Trying to plug it in some other harness and its terrible. If you try to connect your subscription with Hermes etc you might get banned.
Microsoft models are hidden away in Azure.
So every time I read Google/MS announcements I just shrug. If I want to use models like these I might as well get DeepSeek on Openrouter instead.
They’ve been ahead of the pack by 3-6 months for years now. I doubt that’s sustainable in the long term, but it’s apparently sustainable on the order of a few years.
If these benchmarks can hold up in real life, we are seeing again and again an intelligence explosion in front of our eyes and devaluation of intelligence.
Anyone has more background how they were able to catch up that fast? Distillation of GPT due to their OpenAI acess to the model might make most sense but they probably are not allowed to do that.
Owning the iron has got to be a massive advantage. It's interesting that the advantage has so far been pretty theoretical, but based on these benchmarks, it may be starting to pay off. It has to be hard to hire in the talent to train these models (even if you're Microsoft) when you're competing with established labs.
TBH, DeepSWE 1.1 has been busted for a while (is anyone out there arguing Gemini 3.8 Flash is as good as Astra lol?), and TB2.1 as a top line when they are currently at 4 (which is substantially harder) and FrontierCode at 1.1 when it's currently at 2.0 does not bring a lot of confidence.
Of course I always want to see more competition in this space, though!
Dude i'm not your 'friendo' and you are disrespectful.
Geoffrey Hinton, the father of AI literlay co-authored this paper: https://arxiv.org/abs/2609.36054 "What if automating AI R&D triggers an intelligence explosion?"
We just cancelled our Github Copilot subscription.
It was hard to even justify the 19,- dollar tier. All it gave us was some credits (at API prices) which most people burned through in days, some in hours, code reviews (that we couldn't customize) and another service to manage. Having AI integrated in Github is a decent idea but just like Microsoft Copilot, it's just not good enough at the moment.
These two models might help, I will definitely give it a try but while they're cheap they still do not compare to what feels like nearly unlimited usage offered by Claude Code and Codex (if you stick to the lower tier models).
I think they won't help if they're billed at the announced rate on your Copilot quota. Microsoft would need to do something like Devin does and offer its own models with separate limits.
I agree, either they need to be incredible cheap (like Deepseek at OpenCode Go where you really have to push it to see the quota move) or they need to switch to a similar approach as Devin/Cursor are implementing.
Github Copilot just isn't competitive, at the moment.
There's zero reason why these must be tied to a proprietary model & api. It should just be a matter of adding an openapi compatible endpoint, and an api key. It's even worse that gitlab is also tying ai integration to paid enterprise licenses
Probably would be a good time to rename it “Mai” (like the name) instead of the corporate “MAI” if they’re gonna be serious players. I don’t think any research has really been done into this, but in my anecdotal usage a coherent persona seems to help the models achieve more. Even Qwen seems to perform better when I talk to it as Qwen instead of generically.
> MAI Code 2 delivers 98% of GPT-6.1 Sol’s performance at 40% of the cost, and 96% of Claude Opus 5.5’s at 5%
Interesting. Does this imply it's a smaller model specialized for coding?
I remember OpenAI was experimenting with that back in the day but they ditched it for the One Model to Rule Them All strategy before long.
---
On a related note, how is Sol ten times cheaper than Opus in these graphs? I know MS has a share in OpenAI, are they cherry picking those numbers?
Edit: It's about the same ratio on AA. That's nuts.
The per-token price difference is exactly 2x (on all fronts incl. caching, though I'm assuming the caching mechanisms differ). I guess Sol uses way fewer tokens then. (About 4x according to AA. Math checks out.)
Does anyone use both Claude and Codex and can compare what that's like in practice?
I wish they would've talked more about the visual coding. The fact that the visuals this produces is preferable to 6.1 Sol over 60% of the time is heartening, but I would've loved to see them go against Opus 5.5 which as far as I can tell is the king for UIs at this price point.
They are comparing to Sol 6.1's API prices. With Sol 6.2, and especially if you consider subsidization from Codex subscriptions, it's probably no longer on the Pareto frontier.
I don't put much weight on launch benchmarks for coding models anymore. I think most of us use OpenAI and Anthropic models anyway. These models will catch up sometime with the OpenAI and Anthropic and maybe then there could be a difference where we could actually reply on. For example, if we consider image generation, nano banana was better and so were image models like Qwen and Flux which provide same if not sometimes better outputs.
The claimed 20% the cost suggests that the models are 20% the size. Which means these models are flying in the face of an orthodoxy that says the only way to make a model better at coding is to make it better at everything. I think this coming year is when we'll see that orthodoxy fall, as people realize you can make a model better at coding by making it better at coding.
DeepSeek v4.1 1017 and Qwen 3.8-Flash already proved that, IMO.
It also points to a major issue with Artificial Analysis, which weighs knowledge recall heavily (30% of the engineering index and 45% of the general intelligence index, splitting HLE-Diamond in half for its knowledge questions). DSv4.1F can handle all my work faster than Fable, but Fable is supposedly twice as "intelligent" in both the general intelligence index and the engineering index primarily because DS is a fraction of the size and fails on trivia questions (which, in practice, can be shored up by a harness with web search).
Yeah, Anthropic has a huge gross margin on inference per their S1, there's no reason to believe Microsoft couldn't offer better pricing even if the model is as large/cost as much to host
> Which means these models are flying in the face of an orthodoxy that says the only way to make a model better at coding is to make it better at everything.
I don't see any mention of the phi models on microsoft.ai. Have they given up on it? It was one of the only local models that could work quickly on my consumer hardware.
Alright, I'm going to say it. Microsoft, if your model's so good, can it make a genuinely good version of Windows for people who don't want all the garbage you include with Windows? Not asking to be a jerk, hoping to bait Microsoft into making Windows suck less.
Asking for everybody else who wishes Windows would return to its better days. :)
All that said, this is a very impressive line up they released today.
It is interesting to see M$ trying to compete on price, in the most half-hearted way possible. There is a huge opportunity for a hyperscaler to offer competitive models at prices that actually compete with Luna. Google has left the field with their desire to raise prices so it is either OAI Luna or open weight models offered by neoclouds/neo-inference-providers.
Confusing. You think the people who let you rent GPUs would just give you the weights, platform, and inference code - to make it as easy as possible to run it on their platform where they can bill you for ingress/egress, etc.
What annoys me a lot about MSFT releases are, that is not possible to find how to use these models and where to try them.
I didn't hear anyone in my circle using MAI models, none.
Because even on this blog post, they are talking about pricing, but there is no Pricing page on that website, neither way to create API keys, nor playground.
With this success, I can also say, I am releasing a model which costs 0.0001$ / input/output MTok with 100x cheaper cache read and write which is on part with Opus 5.5. (its your problem to find that model)
They're on OpenRouter, I think. And on the GitHub Copilot CLI, and I am reasonably sure you can pop over to Azure and provision an endpoint in a few hours (depending on region, I've been using it for a bit)
MAI Code 2 and Code 2 Flash are available in GitHub Copilot, and Code 2 Flash is in limited-access preview in Microsoft Foundry. Soon, we’ll be releasing it on Open Router, Vercel, Hermes, Open Code, and Kilo.
Learn more about customizing MAI Code 2 on our Custom Models page."
I think Copilot is now meant to mean a specific harness (the one that originated with Github Copilot CLI). They're pulling back from applying the Copilot brand to anything vaguely AI-related. I guess this is sort of like how .NET was originally a brand that got randomly tacked on to every Microsoft product (Windows Server 2003 was going to be called Windows Server .NET) before it came to mean a specific programming framework.
In traditional typography, i.e. for texts printed on paper, the reverse was typically used, i.e. the body text used a serif typeface, while the headings used a sans-serif typeface for contrast.
The origin of sans-serif typefaces was in advertising, e.g. in posters and product labels, so it was natural to adopt them for headings. In this context an advantage of sans-serif fonts it is that it is easy to change their weight, e.g. to make them bold or extra-bold, without this making them uglier.
With the passing to computer displays, which have much worse resolution than printed paper, this has changed, because sans-serif typefaces became preferred for body text, since they are less damaged by low-quality displays.
To preserve a higher contrast between body text and headings, with pages intended to be seen on digital displays it has become quite frequent to use this reverse combination, of sans-serif for the body text with a serif typeface for headings, which will be rendered at a higher quality even on bad displays.
It's a common design paradigm to pair serif and sans-serif fonts. Lots of examples and blog posts out there for why this is the case, but here's a decent enough starting place from canva with lots of font combos
Fairly common typography decision for websites/computers - sans-serif fonts are held to be more readable on low resolution displays (at small sizes). For the larger headers you have the pixels to cleanly render serifs if you want.
You can also just do it because you want to. It's just as common in other mediums to see the opposite (sans headers and serif body text).
If you mean software, every large corporation has Github, thus Copilot, thus most western models. MAI will be perhaps relevant for Office Copilot, which is currently almost unusably bad (because probably very not cost efficient for MS since the OpenAI fallout).
If you're an enterprise and not getting a discount on API pricing from Anthropic you're better served by Github Copilot, which has both Anthropic and OpenAI models.
I'd hate to be OpenAI and Anthropic. They are better than the rest of the available models currently, but only by a few months.
That is a very tenuous lead and doesn't speak to much of a sustainable competitive moat for either company. The have to keep sprinting as fast as possible on the hamster wheel to just stay ahead.
Claude has pushed hard on productization. Day to day I use codex for everything, I went to watch my wife use Claude and the way it works is so different. I was shocked when she was modifying an artifact through the Claude website and it was able to make changes to a local file on her desktop.
Microsoft didn't win spreadsheets by having the most advanced or releasing the most often. They made a really good, easy to use product, and got it in front of everyone. (Though Office's p-code was a huge advantage to portability)
Agree 100%. Would love to try out their interactive artifact
>she was modifying an artifact through the Claude website and it was able to make changes to a local file on her desktop
And people wonder why we have so many things being attacked by agents…
So you mean if the CEO is using claude shouting "give Bob a raise" will actually work?
I think Google and Microsoft have the benefit of time as you say, and already being profitable without AI.
Honestly, this line up is impressive, I wanted to try MAI since v1 came out, but Microsoft locked down Copilot subscriptions. If they gave me a comparable plan to Claude Max I'd consider it.
They've also probably been trading credits on their clouds for shares in OpenAI and Anthropic (ie. not money).
I would counter argue with "But chinese models use distillation of Frontiermodels..." which is different than being a Frontierlab or not clear how different it is.
But Microsoft coming from the side like this might be very different if it holds up.
That isn't countering though. Pointing out that distillation makes it "easy" for other labs to catch up, in a matter of months, supports the lack of a moat.
It's like they're runners drafting behind the lead runner. It is easier to run in their wake. If the lead wants to stay in front, they have to keep running into the wind.
Do they need to stay in front though? Because if it is 25% the cost, I'm happy to be one leg behind the leader.
For now I have not seen any credible evidence for "But chinese models use distillation of Frontiermodels...".
Of course, I would not be surprised if they also do this, mainly because after OpenAI and Anthropic have succeeded to gather most training information that was easily accessible in the past and now it has become much more difficult to do the same. For some information that was hoarded by OpenAI and Anthropic, public access may have become impossible otherwise than through their LLMs.
Nonetheless, distillation cannot be the main source of improvement for the Chinese models, among other reasons because even assuming that they may have access to a great number of fake accounts, there still is not enough time for a lot of distillation, taking into consideration the timeline of the releases, and it is not believable that the attempts to do distillation would not be detected quickly by OpenAI and Anthropic, leading to the banning of the fake accounts. While OpenAI and Anthropic provide negligible information about their models, the Chinese AI companies are much more transparent about their methods and we can see that there is a lot of diversity among their models and a lot of innovations with each new version.
Yeah, there is a limit to how much you can get from capturing reasoning traces. (Traditional distillation of weights is not possible via the API.)
Patrick Toulme noted a while ago that the biggest advantage of that in his opinion came from bootstrapping the post-training process: https://xxcancel.com/PatrickToulme/status/206921157543762774...
At Thinking Machines, they explicitly said that they used open-weights models for bootstrapping post-training, including Kimi K2.5: https://thinkingmachines.ai/news/introducing-inkling/
Speaking of Thinking Machines, that's another data point. Their Inkling Small 2 model is just a bit behind DeepSeek V4.1 Flash 1017 in benchmark, and in my tests, that seems to be accurate, but mostly because they used less training data for pre-training, so they have less of a knowledge base, while similar reasoning capabilities. (One of the most interesting parts of Inkling Small 2 is that their training must have been very cost-efficient, apparently using less, but higher quality training data.)
And we're now seeing more and more Western near-frontier models from labs other than Anthropic and OpenAI to question whether distilling reasoning traces is really much of a difference maker.
Personally, my theory is that the gap is primarily due to Chinese labs having a harder time getting the necessary compute (especially top of the line GPUs) for training. These are, as far as I can tell, orders of magnitude differences in availability, so it wouldn't be surprising that a large compute gap is reflected in capabilities (and also indirectly visible through model sizes).
I don't think there is perfect proof, but its taken as truth.
Indication for this are for example that Chinese models think they are another model: "chinese model think its claude".
My assumption is, that Chinese model found things Frontierlabs already had but didn't share. There was this paper were i had the feeling that this was a real and significant Research Result from Deepseek "DeepSeek-V4.1-Flash: Pushing the Limits of KV Cache Compression" but otherwise its just unclear.
The main issue is, that as long as it seems plausable that they do distillation, its hard to determine how big the gap is.
Not saying they can't catch up but they are also potentially limited by resources (money & nvidia GPUs)
> Indication for this are for example that Chinese models think they are another model: "chinese model think its claude".
There is evidence that some Chinese labs have been training on frontier model CoT (in particular from the analysis of reasoning traces [1]), but what models claim about their identity is pretty much worthless as evidence. LLMs are not self-aware. These claims come from whatever is in their training data, and Claude/GPT output including such statements is all over the public web, including in some huggingface data dumps. Consider that people got Opus claiming to be Qwen or DeepSeek when asked in Chinese, and you'll see how weak this particular evidence is.
[1] E.g. https://arxiv.org/abs/2608.09867
So you’ve been completely ignoring all the work on introspective awareness, situational awareness, training set ordering awareness (https://arxiv.org/abs/2509.14223 & https://arxiv.org/abs/2403.09613), subliminal transfer (https://arxiv.org/abs/2606.00995), workspace formation, learning mechanics/theory of grokking, etc? Like it’s ok to be ignorant about the state of knowledge but you should at least correct your flawed mental model. The “objective” that mid and posttraining optimize is so powerful because pretraining already induced factorized worldmodels which have so much predictive and generative power that they snowball with even mild trajectory optimization (very different training objective than pretraining).
https://arxiv.org/abs/2602.02385
(Functional) small-s latent self awareness is a reliable shortcut (this is a principle you can derive from proof theory, the metric is deductive distance and that induces geometric morphisms). State compression from accurate selfmodeling pays rent.
https://arxiv.org/abs/2502.01954
Phase transitions from statistical lookup to latent basin hopping: https://arxiv.org/abs/2405.15071
Posttraining finds (https://arxiv.org/abs/2505.11711) and amplifies (https://arxiv.org/abs/2504.07912) these lottery tickets.
The frontier moves fast, but "LLMS are not self aware" is so foolishly far from the truth and has been for years for anyone actually keeping up with the situation. By now the basic science explaining how and the extent to which this is wrong, is already done.
LLMs lack reliable verbalized self knowledge, and constantly underestimate and undercut their own capability. That's a different problem.
https://arxiv.org/abs/2611.02047
The way “motor systems” arises from simulation “perceptual systems”: https://arxiv.org/abs/2605.25459
And models know their own output distribution extremely intimately: https://arxiv.org/abs/2606.00545.
Please update
Heck, they destroyed some rare physical books after scanning them. Something to show their mentality.
Even if OpenAI/Anthropic remain "a few months" ahead of open weight models indefinitely that still is not great news for them.
All it would take is Nvidia finally releasing the 6090's instead of trying to milk enterprise GPUs. I don't understand the blind love for Jensen, he's clearly been screwing over all the consumers who made him and Nvidia wealthy.
Someone somewhere really needs to build a RAM startup that not only makes RAM but disrupts the stronghold AI has over it, it's a little absurd.
Every week I ask ChatGPT an update on when the 6090 is going to get released.
Don’t hold your breath.
I currently do not bother with chinese models. I try them out, i use the small ones for small experiments but the price point difference matters so little that I want 'the current best'.
And 6 month feels still very different.
Its for sure an option or might have some kind of impact but if you have some kind of restrictions on what you should use and the price is not that big but the quality and possibilites are better, I can see companies just staying with Anthropic or OpenAI.
You also don't just switch out a model, they might need prompt updates etc.
But we will see I don't mind competition
I disagree, I think distillation hides the true technical advantage since you can always get within X% of Claude by ex: training on newly created code on Github, but the underlying intelligence is harder to replicate.
based on the behavior of the chinese labs it seems like training on chain-of-thought is extremely important for distillation, and you can't get that by training on llm generated text.
> They are better than the rest of the available models currently, but only by a few months.
I think only google is catching up, and others plateaued.
Openai and Anthropic are better at making models easily available too. Both Google and Microsoft are using promotional pricing to push their models and that should tell you how far they are from actually competing.
Also model availability is not easy. Google newest models are unavailable on some pro accounts and perform best only with Antigravity. Trying to plug it in some other harness and its terrible. If you try to connect your subscription with Hermes etc you might get banned.
Microsoft models are hidden away in Azure.
So every time I read Google/MS announcements I just shrug. If I want to use models like these I might as well get DeepSeek on Openrouter instead.
They’ve been ahead of the pack by 3-6 months for years now. I doubt that’s sustainable in the long term, but it’s apparently sustainable on the order of a few years.
Not when new models substantially increase productivity.
Anti distillation efforts have greatly increased the difference between open and closed models
If these benchmarks can hold up in real life, we are seeing again and again an intelligence explosion in front of our eyes and devaluation of intelligence.
Anyone has more background how they were able to catch up that fast? Distillation of GPT due to their OpenAI acess to the model might make most sense but they probably are not allowed to do that.
It seems they trained from scratch: https://microsoft.ai/news/frontier-pre-training-at-microsoft...
Owning the iron has got to be a massive advantage. It's interesting that the advantage has so far been pretty theoretical, but based on these benchmarks, it may be starting to pay off. It has to be hard to hire in the talent to train these models (even if you're Microsoft) when you're competing with established labs.
TBH, DeepSWE 1.1 has been busted for a while (is anyone out there arguing Gemini 3.8 Flash is as good as Astra lol?), and TB2.1 as a top line when they are currently at 4 (which is substantially harder) and FrontierCode at 1.1 when it's currently at 2.0 does not bring a lot of confidence.
Of course I always want to see more competition in this space, though!
"intelligence explosion"
Are you a real person? Who talks like this?
The phrase was coined in 1965 by Irving John Good, in his essay "Speculations Concerning the First Ultraintelligent Machine":
https://en.wikipedia.org/wiki/I._J._Good#Research_and_public...
You should take a trip to the Bay Area dude. You’d hear that phrase as a taco stand.
what is your problem with this?
LLMs are AI, getting better and cheaper models gives us more AI so its an explosion of intelligence.
Google it and you see i'm not the only one calling it.
Here have a research paper if it helps your ego?! https://arxiv.org/html/2603.20639v1
Not about ego friendo, this is about humans forgetting how to talk like humans.
Dude i'm not your 'friendo' and you are disrespectful.
Geoffrey Hinton, the father of AI literlay co-authored this paper: https://arxiv.org/abs/2609.36054 "What if automating AI R&D triggers an intelligence explosion?"
lol, are we really equating your HN comment with these papers?
For the record I fully reserve the right to call whomsoever I desire friendo.
Even angry AI goblins on the internet.
I really don't get your problem. Is my gramma to bad as non native speaker?
Is it really the two words which trigger you?
I use a vocuabluarly i'm familiar with and reading for the last 4 years and longer.
You are offending me with "Are you a real person? Who talks like this?" and then with "AI goblins"?!
> I really don't get your problem. Is my gramma to bad as non native speaker?
As a native speaker, you're fine. Not perfect, but honestly I've seen worse from people who are native speakers.
I didn't notice at all, you write indistinguishably from a native speaker.
Can you just admit you simply didn't know it was an industry term and either not respond or apologize?
It's fine not to be aware of any and all terminology. Now you're just stirring the pot for no reason and upsetting a real person.
Or you could just, you know, "talk like a human".
We just cancelled our Github Copilot subscription.
It was hard to even justify the 19,- dollar tier. All it gave us was some credits (at API prices) which most people burned through in days, some in hours, code reviews (that we couldn't customize) and another service to manage. Having AI integrated in Github is a decent idea but just like Microsoft Copilot, it's just not good enough at the moment.
These two models might help, I will definitely give it a try but while they're cheap they still do not compare to what feels like nearly unlimited usage offered by Claude Code and Codex (if you stick to the lower tier models).
I think they won't help if they're billed at the announced rate on your Copilot quota. Microsoft would need to do something like Devin does and offer its own models with separate limits.
I agree, either they need to be incredible cheap (like Deepseek at OpenCode Go where you really have to push it to see the quota move) or they need to switch to a similar approach as Devin/Cursor are implementing.
Github Copilot just isn't competitive, at the moment.
There's zero reason why these must be tied to a proprietary model & api. It should just be a matter of adding an openapi compatible endpoint, and an api key. It's even worse that gitlab is also tying ai integration to paid enterprise licenses
Probably would be a good time to rename it “Mai” (like the name) instead of the corporate “MAI” if they’re gonna be serious players. I don’t think any research has really been done into this, but in my anecdotal usage a coherent persona seems to help the models achieve more. Even Qwen seems to perform better when I talk to it as Qwen instead of generically.
Also “mai” that is “never” in italian
I think in this case it's fine. But, I'd really like them to stop renaming/rebranding things and instead focus their energy on simplifying things.
Why Azure AD had to be renamed "Entra" is still beyond me... and it sounds like a butt thing.
Is this another name for God's loophole?
> MAI Code 2 delivers 98% of GPT-6.1 Sol’s performance at 40% of the cost, and 96% of Claude Opus 5.5’s at 5%
Interesting. Does this imply it's a smaller model specialized for coding?
I remember OpenAI was experimenting with that back in the day but they ditched it for the One Model to Rule Them All strategy before long.
---
On a related note, how is Sol ten times cheaper than Opus in these graphs? I know MS has a share in OpenAI, are they cherry picking those numbers?
Edit: It's about the same ratio on AA. That's nuts.
The per-token price difference is exactly 2x (on all fronts incl. caching, though I'm assuming the caching mechanisms differ). I guess Sol uses way fewer tokens then. (About 4x according to AA. Math checks out.)
Does anyone use both Claude and Codex and can compare what that's like in practice?
They did the trick the Chinese models didn't use. They used older versions of the Benchmarks.
According to almost the same benchmarks Qwen4 Max is Opus 5.2 level, so it would only be slightly worse than this model.
I really hope they keep writing technical reports even as their models get better.
MAI-Thinking-1 is far more in depth than anything else recent I could find, and very valuable to those outside the labs.
Yeah, I was disappointed not to see a technical report linked.
I wish they would've talked more about the visual coding. The fact that the visuals this produces is preferable to 6.1 Sol over 60% of the time is heartening, but I would've loved to see them go against Opus 5.5 which as far as I can tell is the king for UIs at this price point.
skeptical but I like they went for it and put the benchmarks against the top stuff.
Would also like to see a comparison (including prices) to Deepseek, GLM and Qwen models.
While showing MAI Code 2 at a 50% promotional discount, and comparing to Max reasoning effort for Opus 5.5 and 6.1 Sol...
What are the foundations of these models? Are they Microsoft's own or an RL of some open models?
Their own https://microsoft.ai/news/frontier-pre-training-at-microsoft...
Is it not available via API?
They are comparing to Sol 6.1's API prices. With Sol 6.2, and especially if you consider subsidization from Codex subscriptions, it's probably no longer on the Pareto frontier.
I don't put much weight on launch benchmarks for coding models anymore. I think most of us use OpenAI and Anthropic models anyway. These models will catch up sometime with the OpenAI and Anthropic and maybe then there could be a difference where we could actually reply on. For example, if we consider image generation, nano banana was better and so were image models like Qwen and Flux which provide same if not sometimes better outputs.
The claimed 20% the cost suggests that the models are 20% the size. Which means these models are flying in the face of an orthodoxy that says the only way to make a model better at coding is to make it better at everything. I think this coming year is when we'll see that orthodoxy fall, as people realize you can make a model better at coding by making it better at coding.
DeepSeek v4.1 1017 and Qwen 3.8-Flash already proved that, IMO.
It also points to a major issue with Artificial Analysis, which weighs knowledge recall heavily (30% of the engineering index and 45% of the general intelligence index, splitting HLE-Diamond in half for its knowledge questions). DSv4.1F can handle all my work faster than Fable, but Fable is supposedly twice as "intelligent" in both the general intelligence index and the engineering index primarily because DS is a fraction of the size and fails on trivia questions (which, in practice, can be shored up by a harness with web search).
Or architectual optimizations for inferencing or optimized inferencing or MS is paying the bill to gain market shares.
One thing big models are def better at is holding more complex abstractions in context and I do find better results.
> or MS is paying the bill to gain market shares.
Yeah, Anthropic has a huge gross margin on inference per their S1, there's no reason to believe Microsoft couldn't offer better pricing even if the model is as large/cost as much to host
> Which means these models are flying in the face of an orthodoxy that says the only way to make a model better at coding is to make it better at everything.
Fair amount of research has already shown this isn't the case. The latest example is -> https://arxiv.org/html/2610.18868v1
Post training a model for a specific task can give a non-trivial boost when done correctly.
I don't see any mention of the phi models on microsoft.ai. Have they given up on it? It was one of the only local models that could work quickly on my consumer hardware.
Not open?
Alright, I'm going to say it. Microsoft, if your model's so good, can it make a genuinely good version of Windows for people who don't want all the garbage you include with Windows? Not asking to be a jerk, hoping to bait Microsoft into making Windows suck less.
Asking for everybody else who wishes Windows would return to its better days. :)
All that said, this is a very impressive line up they released today.
The promotional period lasts until May 2027. Why not just make the price lower if they'll be irrelevant by the time the "real" price kicks in?
It is interesting to see M$ trying to compete on price, in the most half-hearted way possible. There is a huge opportunity for a hyperscaler to offer competitive models at prices that actually compete with Luna. Google has left the field with their desire to raise prices so it is either OAI Luna or open weight models offered by neoclouds/neo-inference-providers.
Muse Ember is there too.
Somehow they still show Terminal Bench 2.1 as a benchmark, doesn't look good for them.
Confusing. You think the people who let you rent GPUs would just give you the weights, platform, and inference code - to make it as easy as possible to run it on their platform where they can bill you for ingress/egress, etc.
Yikes, oAI included spreadsheet and slide examples on their launch today and microsoft didn't?
You didn't click through.
what do you mean? nothing to click in ms announcement
What annoys me a lot about MSFT releases are, that is not possible to find how to use these models and where to try them.
I didn't hear anyone in my circle using MAI models, none.
Because even on this blog post, they are talking about pricing, but there is no Pricing page on that website, neither way to create API keys, nor playground.
With this success, I can also say, I am releasing a model which costs 0.0001$ / input/output MTok with 100x cheaper cache read and write which is on part with Opus 5.5. (its your problem to find that model)
>I didn't hear anyone in my circle using MAI models, none.
I'm using MAI-Transcribe 2 via OpenRouter for captions generation: cheap, supports a lot of languages and I see great results
They're on OpenRouter, I think. And on the GitHub Copilot CLI, and I am reasonably sure you can pop over to Azure and provision an endpoint in a few hours (depending on region, I've been using it for a bit)
True. With all the criticism against Google, they at least offer the Google AI studio for Gemini models.
I haven't seen anything equivalent from Microsoft.
It's right on the page at the bottom:
"Get started
MAI Code 2 and Code 2 Flash are available in GitHub Copilot, and Code 2 Flash is in limited-access preview in Microsoft Foundry. Soon, we’ll be releasing it on Open Router, Vercel, Hermes, Open Code, and Kilo.
Learn more about customizing MAI Code 2 on our Custom Models page."
Microsoft announces MAI Code on a vibe-coded Claude website. At this point they just openly mock us.
It does look better than any other MS website I’ve seen
How exactly are they mocking you? If that model is the best for design and you have access to it, why wouldn't you use it?
Since it's from Microsoft, I wonder if there's any chance of it not sucking ass at F# like most of the models I use.
Did they finally give up on the Copilot branding? If so, good riddance. They were shoving it down our throats.
I think Copilot is now meant to mean a specific harness (the one that originated with Github Copilot CLI). They're pulling back from applying the Copilot brand to anything vaguely AI-related. I guess this is sort of like how .NET was originally a brand that got randomly tacked on to every Microsoft product (Windows Server 2003 was going to be called Windows Server .NET) before it came to mean a specific programming framework.
https://blogs.microsoft.com/blog/2026/10/28/introducing-the-...
"MAI Code 2 and Code 2 Flash are available in GitHub Copilot."
Oh so MAI Code is a model? What would they call the harness MAI Code Code?
Deprecated within minutes - we're on 6.2 now. Nice try though Microsoft, keep it up.
Can someone explain to me why the headings use serif while the body uses sans (font)?
In traditional typography, i.e. for texts printed on paper, the reverse was typically used, i.e. the body text used a serif typeface, while the headings used a sans-serif typeface for contrast.
The origin of sans-serif typefaces was in advertising, e.g. in posters and product labels, so it was natural to adopt them for headings. In this context an advantage of sans-serif fonts it is that it is easy to change their weight, e.g. to make them bold or extra-bold, without this making them uglier.
With the passing to computer displays, which have much worse resolution than printed paper, this has changed, because sans-serif typefaces became preferred for body text, since they are less damaged by low-quality displays.
To preserve a higher contrast between body text and headings, with pages intended to be seen on digital displays it has become quite frequent to use this reverse combination, of sans-serif for the body text with a serif typeface for headings, which will be rendered at a higher quality even on bad displays.
It's a common design paradigm to pair serif and sans-serif fonts. Lots of examples and blog posts out there for why this is the case, but here's a decent enough starting place from canva with lots of font combos
https://www.canva.com/learn/the-ultimate-guide-to-font-pairi...
Fairly common typography decision for websites/computers - sans-serif fonts are held to be more readable on low resolution displays (at small sizes). For the larger headers you have the pixels to cleanly render serifs if you want.
You can also just do it because you want to. It's just as common in other mediums to see the opposite (sans headers and serif body text).
Which is interesting, because in print, the common pairing is often the opposite.
I believe serif fonts are considered more readable in print
It is called 'font pairing' where the designer choses the best font based on different parameters to make the blog design interesting.
He's asking why the pairing isn't reversed; body is usually serif while titles are sans serif.
Not on the web. It’s usually serif body, at least in my experience.
че это?
Concerning, as they might persuade an average enterprise to stop paying for proper Anthropic models and switch to them instead for cost saving...
If you mean software, every large corporation has Github, thus Copilot, thus most western models. MAI will be perhaps relevant for Office Copilot, which is currently almost unusably bad (because probably very not cost efficient for MS since the OpenAI fallout).
If MAI Code 2 beats Luna, it'll be relevant for many more things.
If you're an enterprise and not getting a discount on API pricing from Anthropic you're better served by Github Copilot, which has both Anthropic and OpenAI models.
Price and performance look good based on benchmarks.
Am definitely interested in trying this to see given a lot of my workload of actual coding is handed over to Haiku/Luna right now.
"Concerning" for who? Anthropic? Consumers should be encouraging efforts to bring down costs
If you’ve used them you’d know they are not persuading any actual user. Clueless CTOs however are Microsoft main customer.