On Tuesday, November 17, 2026, at its Ignite conference, Microsoft announced two in-house coding models, MAI Code 2 and MAI Code 2 Flash, and GitHub made both available in Copilot the same day. The headline from Microsoft's announcement is a cost comparison: MAI Code 2 costs about 40% of GPT-6.1 Sol and about 5% of Claude Opus 5.5, averaged over four coding benchmarks. The footnote under that claim is the real story. Microsoft says MAI Code 2's cost "is shown at its 50% promotional price." (Sources: Microsoft AI, GitHub changelog)
Quick summary: MAI Code 2 is a general-purpose enterprise coding model and MAI Code 2 Flash is a cheaper model for routine edits, both live in GitHub Copilot as of November 17. Microsoft's own benchmarks put MAI Code 2 at 98% of GPT-6.1 Sol and 96% of Claude Opus 5.5 on a four-test average, at a promotional price that halves every token rate. Per GitHub's pricing docs, that promo runs through May 17, 2027 and the rates then double. The savings are real but smaller at list price, and every performance number is Microsoft's own.
What Microsoft actually shipped
Microsoft describes MAI Code 2 as built for production-ready code, deep debugging, multi-file implementation, repo-wide reasoning and UI work. MAI Code 2 Flash is aimed at high-volume work such as code explanations, targeted edits, routine fixes and rapid iteration. Microsoft is unusually direct about the limits: the models are "explicitly not yet tuned for harder problems, like fundamental science research and very long-horizon agentic work." In plain terms, this is a workhorse pitch, not a bid to top every leaderboard. (Source: Microsoft AI)
- GitHub Copilot: MAI Code 2 is listed for Pro, Pro+, Max, Business and Enterprise plans. MAI Code 2 Flash is also listed for Free and Student plans, per GitHub's changelog.
- Where it runs: GitHub lists Visual Studio Code, Visual Studio, Copilot CLI, the cloud agent, the Copilot app, github.com, GitHub Mobile, JetBrains, Xcode and Eclipse.
- Billing: both models are billed under usage-based billing, not as a flat included allowance.
- Microsoft Foundry: MAI Code 2 Flash is in a limited-access preview there. Microsoft's page does not state a Foundry status for MAI Code 2.
- Coming soon: Microsoft lists OpenRouter, Vercel, Hermes, OpenCode and Kilo as upcoming availability.
- Admin default: GitHub says Business and Enterprise admins manage access through the Copilot model policy, and new models are enabled by default unless an admin turns off the global default or disables the model.
The benchmark claim, with the numbers
Microsoft's headline performance figure is an average of four coding tests: DeepSWE 1.1, Terminal Bench 2.1, SWE-Atlas and FrontierCode v1.1. On that average, Microsoft says MAI Code 2 reaches 98% of GPT-6.1 Sol and 96% of Claude Opus 5.5. InfoWorld tabulated the underlying scores from Microsoft's materials, and they show MAI Code 2 behind both rivals on the average, not ahead of them. Microsoft ran the coding tests on its internal MSBench harness with the GitHub Copilot CLI, without subagents, and all models at maximum reasoning effort. The FrontierCode scores and costs come from Artificial Analysis. (Sources: Microsoft AI, InfoWorld)
| Model | DeepSWE 1.1 / Terminal Bench 2.1 | SWE-Atlas / FrontierCode v1.1 | Four-test average |
|---|---|---|---|
| MAI Code 2 | 73.2% / 84.6% | 43.1% / 52% | 63.2% |
| GPT-6.1 Sol | 74.6% / 85% | 47.4% / 49.4% | 64.1% |
| Claude Opus 5.5 | 69% / 78.7% | 57.5% / 56% | 65.3% |
Read that table for shape, not just totals. MAI Code 2 beats Opus 5.5 on DeepSWE and Terminal Bench but is well behind on SWE-Atlas, 43.1% to 57.5%. An average hides that kind of split, which is why a single "96% of" figure should not decide which model you pick for your own repository. Microsoft also says developers preferred MAI Code 2 over GPT-6 Sol 64% of the time in blind side-by-side visual frontend tests with 258 annotators. That is a preference test on UI output, and it says little about debugging or refactoring. (Sources: Microsoft AI, InfoWorld)
The price math: promo versus list
GitHub's models and pricing documentation lists MAI Code 2 at $0.50 per million input tokens and $2.50 per million output tokens as promotional pricing through May 17, 2027, then $1.00 and $5.00. MAI Code 2 Flash is listed at $0.05 and $0.25 on promo, then $0.10 and $0.50. The same page lists GPT-6.1 Sol at $2 input and $10 output (for contexts up to 272K tokens; higher beyond that) and Claude Opus 5.5 at $4 input and $20 output. InfoWorld reports the same MAI figures. The GitHub changelog itself only says the models are billed under usage-based billing, so check the pricing page again before budgeting. (Sources: GitHub docs, InfoWorld, Beri.net)
| Model (Copilot, per 1M tokens) | Input / output | Cost of 1M input + 100K output |
|---|---|---|
| MAI Code 2, promo | $0.50 / $2.50 | $0.75 |
| MAI Code 2, after May 17, 2027 | $1.00 / $5.00 | $1.50 |
| GPT-6.1 Sol | $2.00 / $10.00 | $3.00 |
| Claude Opus 5.5 | $4.00 / $20.00 | $6.00 |
The last column is our own arithmetic on those reported rates for an illustrative task, not a Microsoft figure. On raw token prices, promo MAI Code 2 costs 12.5% of Opus 5.5 ($0.75 against $6.00), and list-price MAI Code 2 costs 25%. Microsoft's 5% figure is lower than either because it measures cost per benchmark task, which also reflects how many tokens each model burns to finish. That is a fair way to compare, but it is Microsoft's measurement on Microsoft's harness, and it uses the promotional price. Beri.net's analysis concludes that at list price the cost ratios roughly double, and that MAI Code 2 Flash comes out near 130% of GPT-6 Luna's per-task cost instead of the roughly 65% implied by promo pricing. Treat that as one analyst's calculation, but the direction follows directly from the footnote.
Put a calendar reminder for May 2027. If a team standardizes on a model because of a promo price, the budget case has a built-in expiry date.
What Flash is for, and what it is not
For MAI Code 2 Flash, Microsoft reports a three-test comparison (DeepSWE 1.1, Terminal Bench 2.1 and SWE-Atlas) at maximum reasoning. InfoWorld's table lists Flash at an average of 61%, against 55.1% for GPT-6 Luna and 53.9% for Claude Haiku 5.5. Unlike MAI Code 2, Flash leads its comparison set in Microsoft's own numbers. The mid-sized lesson is that Flash is built for the long tail of boring coding requests: explain this function, rename this variable, fix this lint error. Those are the requests that quietly add up under usage-based billing. (Sources: Microsoft AI, InfoWorld)
Microsoft also says a Flash-derived model, MAI Work Flash, powers Microsoft 365 and Fabric features. It reports parity with GPT-6 Luna in Excel and PowerPoint on internal benchmarks, and in Word similar results to GPT-6.1 Sol using about 42% fewer input tokens per call in production. Those are internal evaluations that readers cannot reproduce, so they belong in the "interesting, unverified" pile. One housekeeping note from the same Beri.net write-up: the older MAI-Code-1.1-Flash is reported to retire across all Copilot experiences on December 18, 2026, which would break scripts that name it explicitly. Confirm that against GitHub's own notice before acting on it.
A 10-pull-request test you can run this week
Vendor benchmarks run on vendor harnesses. Your repository has its own build system, test suite and conventions, so the only comparison that matters is one you run yourself. A small, cheap version:
- Pick 10 closed pull requests from the last quarter, mixing bug fixes, small features and one refactor. Keep the original issue text and the tests.
- Run each task with MAI Code 2, MAI Code 2 Flash and the model you use today, with the same instructions and the same reasoning setting.
- Record cost from the usage dashboard, not from the model's token count alone. Note any retries.
- Score on whether tests pass, whether a reviewer would merge it with minor edits, and how long review took.
- Compute cost per merged change, at both promo and list price, since list price is where you will live after May 17, 2027.
- Check that the result survives the weakest category. If Flash handles the five routine tasks at passing quality, route routine work to it and keep the heavier model for the rest.
What to check before you flip it on
- Model policy: new models are on by default for Business and Enterprise unless an admin changes the default, so decide on purpose instead of by accident.
- Billing mode: both models bill by usage. Set budget alerts before a team of developers discovers agent mode.
- Task fit: Microsoft itself says very long-horizon agentic work is not the target. Do not benchmark it on that and call it a failure.
- Data and security review: model cards and compliance terms matter more than a price chart. Read them as you would for any new model in your toolchain.
- Independent evidence: as of November 17, the performance numbers we found are Microsoft's, and only FrontierCode comes from a third party (Artificial Analysis).
01Is MAI Code 2 available to free GitHub Copilot users?
According to GitHub's changelog, MAI Code 2 is listed for Pro, Pro+, Max, Business and Enterprise. MAI Code 2 Flash is also listed for Free and Student plans. Both are available to eligible subscribers on annual plans.
02How much does MAI Code 2 cost?
Per GitHub's pricing documentation, $0.50 per million input tokens and $2.50 per million output tokens until May 17, 2027, then $1.00 and $5.00. Verify against GitHub's current pricing documentation before budgeting, and note it is billed per usage.
03Is MAI Code 2 better than Claude Opus 5.5 or GPT-6.1 Sol?
On Microsoft's four-test average, it scores slightly lower than both (98% of GPT-6.1 Sol and 96% of Opus 5.5, per Microsoft). It wins some tests and loses others. The case for it is cost and everyday coding, not a higher ceiling.
04Is the 5% cost claim misleading?
It is accurate to Microsoft's footnote but conditional. It uses the 50% promotional price and Microsoft's own cost-per-task measurement. At list price the gap shrinks, so run your own test.
05Can I use MAI Code 2 outside GitHub Copilot?
Microsoft lists OpenRouter, Vercel, Hermes, OpenCode and Kilo as coming soon, and Foundry has a limited preview of MAI Code 2 Flash. Dates were not given.
Our take
A coding model that lands within a few percent of the leaders at a fraction of the token price is useful news for anyone running agents all day, and Microsoft deserves credit for stating what the model is not tuned for. The sensible reaction is neither hype nor cynicism. Treat the 5% number as a ceiling that applies until May 2027, treat the benchmarks as an invitation to test, and pick the model on cost per merged change in your own code.
If you want a quiet place to compare how different AI models handle the same coding prompt before you commit a team's budget to one of them, LumiChats lets you put them side by side in one chat. It is a handy way to sanity-check a vendor's claims with your own question first.
