Dirty Static // The Verdict
Near frontier scores on everyday coding in Microsoft own tests, at a small fraction of the cost per task of the models it trails. The catch is in the fine print: the tests are Microsoft tests, and the price that makes the maths sing doubles in May 2027.
Best forCopilot teams doing everyday enterprise coding, deep debugging, changes across many files and front end work, who want a strong model on the Pro plan without paying frontier rates
What works
- Averages 63.2 percent across four coding benchmarks in Microsoft tests, close to GPT-6.1 Sol at 64.1 percent and Claude Opus 5.5 at 65.3 percent
- Average cost per task of $0.309 in those tests, against $0.761 for Sol and $6.306 for Opus 5.5
- Token rates of $0.50 input and $2.50 output per million tokens in Copilot until May 17, 2027, a quarter of the Sol rate
- A 1,050,000 token context window with text and image input, five reasoning levels, and a place on every paid Copilot plan including Pro
What does not
- Every headline score and cost figure comes from Microsoft, with three of four benchmarks run on its own platform and no independent result at launch
- The cost comparison uses a 50 percent promotional price, and every token rate doubles on May 17, 2027
- Trails Opus 5.5 by more than fourteen points on SWE Atlas, and Microsoft says it is not yet tuned for very long horizon agent work
- The large model lives only inside GitHub Copilot at launch, with no Foundry deployment and no general API route
Research review: based on published testing, specifications, and documentation, not hands on testing.
Microsoft did not try to win the frontier race on November 17. It tried to win the invoice. MAI Code 2, launched at Ignite in San Francisco with a smaller sibling called MAI Code 2 Flash, scores a little below GPT-6.1 Sol and Claude Opus 5.5 in Microsoft own coding tests and costs a lot less per finished task than either of them. In a market where the best model changes every few weeks, that is a calmer and more useful pitch than another point on a leaderboard.
The pitch also needs reading carefully, because both halves of it carry conditions. The scores are Microsoft scores, run mostly on Microsoft own platform. The cost lead rests on a promotional price with a published end date. Neither condition makes the model uninteresting. Both decide how much weight the launch numbers can carry, and this review weighs them in that order.
What It Is
MAI Code 2 is the larger model in a two model family from Microsoft AI, the in house lab led by Mustafa Suleyman. Both models were trained by Microsoft from scratch and announced on November 17, 2026 as part of a set of seven new MAI models at Ignite. MAI Code 2 is built for complex work in real repositories. MAI Code 2 Flash is the smaller and faster model for everyday work such as code explanations, targeted edits and routine fixes.
Microsoft is unusually plain about the boundary. The company says the pair were built for the coding work its customers do every day and are explicitly not yet tuned for harder problems, naming fundamental science research and very long horizon agentic work. That sentence does more for the credibility of this launch than the benchmark chart does. A vendor that tells you where its model stops is easier to trust on where it starts.
Features and Specifications
| Area | MAI Code 2 specification |
|---|---|
| Maker | Microsoft AI, Microsoft Corporation |
| Released | November 17, 2026, at Microsoft Ignite |
| Architecture | Decoder only transformer with sparse Mixture of Experts layers and an integrated vision encoder |
| Size | About 2.8 trillion total parameters, 77 billion active, plus 3 billion embedded, per the model card |
| Training | About 50 trillion tokens of pre training, plus about 150 billion tokens for long context extension, August to November 2026 |
| Context window | 1,050,000 tokens, with up to 922,000 input tokens and 128,000 output tokens |
| Input and output | Text and image in, text out |
| Reasoning control | Five effort levels, from low to max |
| Knowledge cutoff | August 2026, per the model card |
| Languages | Optimised for English code production only at release |
| Sibling model | MAI Code 2 Flash, about 398 billion total parameters and 12 billion active |
Pricing in GitHub Copilot, per million tokens, checked November 19, 2026
- MAI Code 2 input, to May 17, 2027
- $0.50
- MAI Code 2 cached input, to May 17, 2027
- $0.025
- MAI Code 2 output, to May 17, 2027
- $2.50
- MAI Code 2 input, from May 17, 2027
- $1.00
- MAI Code 2 output, from May 17, 2027
- $5.00
- MAI Code 2 Flash input, to May 17, 2027
- $0.05
- MAI Code 2 Flash output, to May 17, 2027
- $0.25
- Claude Opus 5.5 input and output
- $4.00 and $20.00
- GPT-6.1 Sol input and output
- $2.00 and $10.00
Access is part of the specification here, because the large model has exactly one home at launch. MAI Code 2 sits in the Copilot model picker on the Pro, Pro+, Max, Business and Enterprise plans, and GitHub says eligible subscribers on annual plans get it too. Flash reaches further down, onto Free and Student plans, though those plans reach models only through auto model selection. Flash is also the only one of the pair with an API story at launch, a limited access preview in Microsoft Foundry through an OpenAI compatible Responses API. Microsoft says wider routes through services including OpenRouter, Vercel, Hermes, OpenCode and Kilo are coming, with no date given. Readers following the routing question will recognise the pattern from our look at OpenRouter Auto Code and automatic model choice: the cheaper the model, the more its value depends on how easily tools can reach it.

Real World Experience
This is a research review. Dirty Static did not run MAI Code 2, and no hands on testing is claimed here. What follows draws on the Microsoft announcement, the 24 page model card dated November 17, 2026, GitHub pricing and plan documentation, and launch coverage from DevOps.com and other outlets, all checked on November 19, 2026. At the time of checking, no tester outside Microsoft had published results for this model.
Start with the table behind the headline, because the headline is an average and averages hide the interesting rows. Microsoft averaged four benchmarks: DeepSWE 1.1, Terminal-Bench 2.1, SWE Atlas and FrontierCode v1.1. Three ran on MSBench, Microsoft internal evaluation platform, using the GitHub Copilot CLI at version 1.0.102 with subagents disabled and every model at maximum reasoning effort. The FrontierCode scores and costs came from Artificial Analysis. Results are mean single attempt pass rates, averaged over three runs, with infrastructure errors counted as failures.
| Benchmark | MAI Code 2 | GPT-6.1 Sol | Claude Opus 5.5 |
|---|---|---|---|
| DeepSWE 1.1 | 73.2% | 74.6% | 69.0% |
| Terminal-Bench 2.1 | 84.6% | 85.0% | 78.7% |
| SWE Atlas | 43.1% | 47.4% | 57.5% |
| FrontierCode v1.1 | 52.0% | 49.4% | 56.0% |
| Average | 63.2% | 64.1% | 65.3% |
| Average cost per task | $0.309 | $0.761 | $6.306 |
Read row by row and a more honest picture appears. On DeepSWE, MAI Code 2 beats Opus 5.5 and sits a point behind Sol. On Terminal-Bench it nearly matches Sol and clears Opus 5.5 by almost six points. On FrontierCode it beats Sol and trails Opus 5.5. Then SWE Atlas reverses the story: at 43.1 percent it trails Sol by more than four points and Opus 5.5 by more than fourteen. SWE Atlas covers codebase question answering, test writing and refactoring across 18 maintained open source repositories, which is close to the daily work this model is sold for. That one row is the strongest reason to trial the model on your own repository before routing real work to it.
Average pass rate across four coding benchmarks, percent (Microsoft tests)
DeepSWE 1.1, Terminal-Bench 2.1, SWE Atlas and FrontierCode v1.1, at maximum reasoning effort. Three benchmarks run on Microsoft MSBench platform; FrontierCode figures from Artificial Analysis. Company reported results, not measured by this magazine.
The cost column is where the model earns its headline, and where the footnote lives. An average task costs $0.309 on MAI Code 2, $0.761 on Sol and $6.306 on Opus 5.5 in these tests. Microsoft states that the MAI Code 2 cost is shown at its 50 percent promotional price. Double that price, as GitHub will on May 17, 2027, and the same token use would put the same task near $0.62, still below Sol on this table and near a tenth of Opus 5.5, but no longer at 40 percent and 5 percent. That recalculation is simple arithmetic on Microsoft figures, not a new measurement, and it is the version a budget should be built on. Our review of Gemini 4 Carbon and its December price step made the same point from the other side: a launch price with an expiry date is a promotion, not a rate card.
Flash deserves its own paragraph, because for many teams it will be the model they actually meet. Across DeepSWE 1.1, Terminal-Bench 2.1 and SWE Atlas, Microsoft reports Flash at a 61.0 percent average and $0.052 per task, against GPT-6 Luna at 55.1 percent and $0.08 and Claude Haiku 5.5 at 53.9 percent and $0.319, again at promotional prices. Flash even edges its larger sibling on SWE Atlas, 44.8 percent against 43.1 percent, a reminder that the smaller model was tuned hard for routine work. At a tenth of the Code 2 token rate, Flash is the natural default for explanations, small edits and fixes, with the large model kept for the tasks that earn it.
One more claim sits beside the pass rates, and it is about taste rather than correctness. Microsoft says that in blind side by side evaluations of front end work, covering interfaces, games, data visualisations, 3D scenes and SVG graphics, reviewers preferred MAI Code 2 output over GPT-6 Sol output 64 percent of the time, across 258 annotators. Preference is not a pass rate, and confident looking interface code can win a comparison it would fail in production. Still, front end work is a large share of what Copilot users ask for, and a cheap model that people prefer looking at will get used.
The model card adds the limits that launch posts tend to skip. In long agent sessions the models can repeat tool calls, make too little use of the existing codebase, drift from instructions or stop before fully verifying their work, and quality may fall as context nears its maximum length. Higher reasoning effort raises latency and token use and does not always improve results. Generated code may be incorrect, incomplete, insecure or out of date, and the knowledge cutoff is August 2026. GitHub adds a product level warning of its own: MAI models are continuously improving models, so behaviour may change as new checkpoints arrive, without a new name to mark the change.
What Works
The price performance case is real, even after the promotional discount is stripped out. Token rates are published, checkable and low: a quarter of Sol and an eighth of Opus 5.5 per token at launch rates, and half of Sol and a quarter of Opus 5.5 after May 2027. A model does not need to win a benchmark for that to matter. It needs to be close enough that the saving buys more attempts, more review and more coverage of the boring tasks that fill a sprint.
Placement on the Pro plan is the second genuine win. At $10 a month, Pro is where individual developers and small teams actually live, and GitHub plan table gives Pro no access to Opus 5.5 or Sol as selectable models. MAI Code 2 gives that tier a powerful class model of its own, inside the picker those users already have. That is a distribution advantage no API only launch can match, and it is the same instinct behind the wider Copilot push we covered in Copilot Autopilot and the always on agent.
The honesty about scope is the third win, and it is worth naming because it is rare. Microsoft says in its own launch material that these models are not yet tuned for fundamental science research or very long horizon agent work, publishes a model card with known limitations in plain language, and labels the promotional basis of its cost chart. A buyer can plan around a launch like that.

What Does Not
The evidence problem comes first. Every performance figure in this review traces back to Microsoft: its announcement, its model card, its internal platform, its choice of benchmarks and its choice of rivals, with one of four benchmarks supplied by Artificial Analysis. There is no SWE-bench score in the launch material and no independent leaderboard row to set beside the company table. None of that makes the numbers wrong. It makes them a starting claim, and the gap on SWE Atlas shows why the distinction matters.
The promotional price is the second problem, because the launch story will outlive the promotion. The 98 percent and 40 percent pairing will be quoted for months, usually without the footnote that the cost side was calculated at half price. After May 17, 2027 the token rates double and the task cost gap narrows a lot against Sol, the rival most buyers would otherwise choose. Plan at the May rates and treat the first six months as a discount period, because that is what GitHub documentation says they are.
The third problem is reach. MAI Code 2 itself has no Foundry deployment at launch, no weights to run and no general API to call from your own pipeline or agent framework. If your tooling lives outside Copilot, the model you can actually evaluate today is Flash in a limited access preview. Microsoft promises more routes soon. Soon is not a date, and roadmaps are not distribution.
Competition
| Model | Average pass rate, Microsoft table | Average cost per task | Token price in Copilot, per million | The trade |
|---|---|---|---|---|
| MAI Code 2 | 63.2% | $0.309 at promotional price | $0.50 input, $2.50 output, doubling May 17, 2027 | Near frontier scores in its maker tests, inside Copilot only, at the lowest task cost in this table |
| GPT-6.1 Sol | 64.1% | $0.761 | $2.00 input, $10.00 output | A point ahead on average and ahead on SWE Atlas, at about two and a half times the task cost |
| Claude Opus 5.5 | 65.3% | $6.306 | $4.00 input, $20.00 output | The average leader and the clear SWE Atlas winner, at about twenty times the task cost |
| MAI Code 2 Flash | 61.0% across three benchmarks | $0.052 at promotional price | $0.05 input, $0.25 output, doubling May 17, 2027 | A tenth of the Code 2 token rate, on every plan, and the only family member with a Foundry preview |
| Claude Haiku 5.5 | 53.9% across three benchmarks, per Microsoft | $0.319 | $0.10 input, $0.50 output | The small model rival Flash beats on both score and task cost in Microsoft figures |
The table points to a routing answer rather than a single winner. Flash for volume work, Code 2 for the harder daily tasks, and a frontier model kept in reserve for the long and difficult jobs where SWE Atlas style depth decides the result. That is less exciting than a coronation. It is also how Copilot bills, by the token and by the task, so it is the answer the pricing rewards.
Final Verdict
Try MAI Code 2 if you pay for Copilot, especially on Pro, and your week is full of debugging, refactoring, changes across many files and front end work. Pick it in the model picker, run it on tasks you know well, and compare the results and the credit use against the model you use now. On Microsoft evidence it will land close to the frontier models on most of that work and cost far less doing it, and your own repository is the only test that can confirm the claim where it counts.
Keep a frontier model for the work Microsoft itself flags, the very long agent runs and the deep repository reasoning where Opus 5.5 holds a clear lead in the one benchmark that measures it best. And write May 17, 2027 in the budget now, while the promotional price is still the one on the invoice. MAI Code 2 does not need to be the best coding model in the world to be worth using. It needs to be good enough, often enough, at this price, and on the published evidence it clears that bar with room to spare. The evidence is one company deep, so verify it on your own code before you build on it.
Sources
- Microsoft AI: MAI Code 2 and MAI Code 2 Flash model card, November 17, 2026
- Aistify: Microsoft launches MAI Code 2 and Flash for cheaper AI coding
- DevOps.com: At Ignite, Microsoft makes the case for good enough coding models
- Enjoy.dev: MAI Code 2 is in GitHub Copilot at an eighth of Opus 5.5 price
- OrcaRouter: Seven MAI models at Ignite 2026, where MAI Code 2 fits in
- Images: Hero, repository and scale images are generated Dirty Static house style illustrations.
END OF TRANSMISSION
Stay curious. Question everything.


