Microsoft launched MAI Code 2 at Ignite on 17 November 2026, and its pitch is price. Microsoft says the new coding model reaches 98% of GPT-6.1 Sol's performance at 40% of the cost, and 96% of Claude Opus 5.5's at 5% of the cost. A smaller sibling, MAI Code 2 Flash, targets quick edits and routine fixes. Both models are live in GitHub Copilot today.
The headline ratios carry a footnote most coverage skipped. Microsoft priced its own models at a 50% promotional rate for the comparison, and GitHub's pricing page says that rate ends on 17 May 2027. This article covers what MAI Code 2 is, the full benchmark results including the test it loses, and what it costs before and after the promotion. It also covers which Copilot plans get it and what the rest of Microsoft's seven-model Ignite lineup adds.
The Key Takeaways
- What it is: MAI Code 2 is Microsoft AI's in-house coding model for everyday enterprise work, launched at Ignite on 17 November 2026 with a cheaper MAI Code 2 Flash.
- The claim: 63.2% average pass rate across four coding benchmarks, against 64.1% for GPT-6.1 Sol and 65.3% for Claude Opus 5.5, at $0.309 per task against $0.761 and $6.306.
- The catch: Microsoft's cost figure uses a 50% promo price. At the list price from 17 May 2027, MAI Code 2 costs about 81% of GPT-6.1 Sol per task, not 40%.
- Price: $0.50 input and $2.50 output per million tokens until 17 May 2027, then $1 and $5.
- Where to get it: GitHub Copilot Pro, Pro+, Max, Business and Enterprise. Code 2 Flash also reaches Free and Student, and MAI-Code-1.1-Flash retires on 18 December 2026.
What Is MAI Code 2?
MAI Code 2 is a general-purpose coding model built by Microsoft AI, the in-house lab run by Mustafa Suleyman. Microsoft says it is tuned for "the everyday enterprise coding tasks" it sees in GitHub Copilot. Its announcement lists taking code to production, deep debugging, multi-file implementation, repository-wide reasoning and UI work.
MAI Code 2 is the second generation of Microsoft's coding line. The first model, MAI-Code-1-Flash, arrived at Build in June as part of Microsoft's first seven MAI models. MAI-Code-1.1-Flash followed in August. Both were small, fast models. MAI Code 2 is not: GitHub files it under its "Powerful" category, the same bracket as GPT-6.1 Sol and Claude Opus 5.5, while the Flash models sit under "Lightweight".
Microsoft is upfront about the limits. Its announcement says both new models are "explicitly not yet tuned for harder problems, like fundamental science research and very long horizon agentic work." That caveat matters for anyone who runs Copilot's cloud agent on hour-long tasks.
MAI Code 2 vs MAI Code 2 Flash
MAI Code 2 Flash is the lightweight, high-volume version. GitHub describes it as suited to "code explanations, targeted edits, routine fixes, and rapid iteration." The bigger model is the one for multi-file changes and debugging. Flash costs one tenth as much per token, and it is the only one of the pair that free Copilot users get.
Microsoft compares the two models against different rivals. The full model is measured against GPT-6.1 Sol and Claude Opus 5.5. Flash is measured against GPT-6 Luna and Claude Haiku 5.5.
Microsoft AI's own launch post on X leads with price, not with capability:
Frontier performance. A fraction of today’s prices. Built to run 24/7.
— Microsoft AI (@MicrosoftAI) November 17, 2026
Meet the new MAI models, launching today at #MSIgnite: https://aka.ms/MAI-Ignite-2026
MAI Code 2 Benchmarks: What Microsoft Claims
MAI Code 2 scores 63.2% on average across four coding benchmarks in Microsoft's tests. That is 0.9 points behind GPT-6.1 Sol and 2.1 points behind Claude Opus 5.5. The four benchmarks are DeepSWE 1.1, Terminal-Bench 2.1, SWE Atlas and FrontierCode v1.1, with every model run at maximum reasoning effort.
Microsoft AI calls these four "a basket of practical, everyday coding benchmarks that best reflect the usage we see in GitHub Copilot."
The averages below come from the chart in Microsoft's MAI Code 2 announcement. Cost per task is the mean cost to complete one benchmark task, in US dollars.
| Model | Average pass rate | Average cost per task |
|---|---|---|
| MAI Code 2 (promo price) | 63.2% | $0.309 |
| GPT-6.1 Sol | 64.1% | $0.761 |
| Claude Opus 5.5 | 65.3% | $6.306 |
| MAI Code 2 Flash (promo price) | 61.0% | $0.052 |
| GPT-6 Luna | 55.1% | $0.080 |
| Claude Haiku 5.5 | 53.9% | $0.319 |
The Flash rows use three benchmarks, not four, because FrontierCode is left out. On those numbers, MAI Code 2 Flash is not just "comparable" to GPT-6 Luna and Claude Haiku 5.5, as Microsoft puts it. It scores six to seven points higher than both.
MAI Code 2 Results by Benchmark
MAI Code 2 does not win any of the four benchmarks outright, and it loses one clearly. The averages hide that spread, so here are the individual scores from the same chart:
| 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% |
Where MAI Code 2 Falls Short
SWE Atlas is the weak spot. Microsoft's model scores 43.1% there, 14.4 points behind Claude Opus 5.5. Opus loses to both rivals on DeepSWE and Terminal-Bench but wins SWE Atlas and FrontierCode, which is why it tops the average. MAI Code 2 beats GPT-6.1 Sol only on FrontierCode, by 2.6 points.
Three conditions limit how far these numbers travel. Microsoft ran DeepSWE, Terminal-Bench and SWE Atlas on MSBench, its internal evaluation platform, through the GitHub Copilot CLI without subagents. Only the FrontierCode scores come from Artificial Analysis, an outside lab. And none of these numbers had been independently reproduced at the time of writing.
Microsoft also ran a blind test of frontend work: interfaces, games, data visualizations, 3D scenes and SVG graphics. Its 258 annotators preferred MAI Code 2 over GPT-6 Sol 64% of the time. Some early coverage reported 67%, but Microsoft's own post says 64%.
MAI Code 2 Pricing: The Promo Does the Heavy Lifting
MAI Code 2 costs $0.50 per million input tokens and $2.50 per million output tokens in GitHub Copilot until 17 May 2027. From that date, the price doubles to $1 and $5. Those figures come from GitHub's Copilot models and pricing page, which bills both MAI models under usage-based billing.
| Model (per 1M tokens) | Input | Cached input | Output |
|---|---|---|---|
| MAI Code 2, promo to 17 May 2027 | $0.50 | $0.025 | $2.50 |
| MAI Code 2, from 17 May 2027 | $1.00 | $0.05 | $5.00 |
| MAI Code 2 Flash, promo to 17 May 2027 | $0.05 | $0.005 | $0.25 |
| MAI Code 2 Flash, from 17 May 2027 | $0.10 | $0.01 | $0.50 |
| GPT-6.1 Sol (up to 272K) | $2.00 | $0.10 | $10.00 |
| Claude Opus 5.5 | $4.00 | $0.20 | $20.00 |
Microsoft's Visual Studio Code team announced the discount alongside the launch:
🚀 MAI Code 2 and MAI Code 2 Flash are now available in VS Code with GitHub Copilot!
— Visual Studio Code (@code) November 17, 2026
From debugging and building features across multiple files to quick edits and everyday fixes, choose the model that fits your coding task.
⚡️ Enjoy 50% off both models through May 17, 2027!
Try them today!
🔗 https://github.blog/changelog/2026-11-17-mai-code-2-and-mai-code-2-flash-in-github-copilot/
Microsoft's footnote is plain about the comparison: "MAI-Code-2 cost is shown at its 50% promotional price." The same note sits under the Flash results. So the "40% of the cost" headline holds for six months, not for good.
MAI Code 2 Cost at List Price
At list price, MAI Code 2 costs about 81% of GPT-6.1 Sol per task. The promotion halves every token class (input, cached input and output), so the list-price cost per task is exactly double Microsoft's figure. We did that arithmetic on Microsoft's own chart data:
| Comparison | Promo cost per task | List cost per task | Share of rival, promo | Share of rival, list |
|---|---|---|---|---|
| MAI Code 2 vs GPT-6.1 Sol | $0.309 vs $0.761 | $0.618 vs $0.761 | 41% | 81% |
| MAI Code 2 vs Claude Opus 5.5 | $0.309 vs $6.306 | $0.618 vs $6.306 | 5% | 10% |
| Code 2 Flash vs GPT-6 Luna | $0.052 vs $0.080 | $0.104 vs $0.080 | 65% | 130% |
| Code 2 Flash vs Claude Haiku 5.5 | $0.052 vs $0.319 | $0.104 vs $0.319 | 16% | 33% |
Two results stand out. Against Claude Opus 5.5, Microsoft's model stays dramatically cheaper either way, at a tenth of the cost per task for 96.8% of the score. Against GPT-6.1 Sol the saving shrinks to about a fifth after May. MAI Code 2 Flash ends up 30% more expensive per task than GPT-6 Luna at list price, though it still scores higher.
Microsoft's comparison has also aged in one day. OpenAI shipped GPT-6.2 on the same 17 November, and GitHub lists GPT-6.2 Sol at the same $2 and $10 as GPT-6.1 Sol. Microsoft's charts compare against the older GPT-6.1 Sol.
How to Use MAI Code 2 in GitHub Copilot
MAI Code 2 is available in GitHub Copilot on the Pro, Pro+, Max, Business and Enterprise plans. MAI Code 2 Flash reaches the same plans plus Copilot Free and Student. GitHub's launch changelog says both models are also open to eligible subscribers on annual plans.
You pick either model from the Copilot model picker. GitHub lists Visual Studio Code, Visual Studio, Copilot CLI, the Copilot cloud agent, the GitHub Copilot app, github.com, GitHub Mobile on iOS and Android, JetBrains IDEs, Xcode and Eclipse. If you are not paying for Copilot yet, our guide on how to get GitHub Copilot Pro free covers the student and open-source routes.
On Copilot Business and Enterprise, administrators control MAI models through the model policy in Copilot settings. New models are on by default unless an admin has turned off the global default or blocked the model.
Existing users of the older Flash model have a deadline. GitHub will deprecate MAI-Code-1.1-Flash on 18 December 2026 across every Copilot experience, according to its deprecation notice, and points users to the latest MAI Code Flash model. Enterprise admins may need to enable MAI Code 2 Flash in their model policy first.
MAI Work Flash: The Same Model Inside Microsoft 365
MAI Work Flash is a version of MAI Code 2 Flash that Microsoft tuned further to serve Excel, PowerPoint, Word, Teams and Fabric IQ. It shows the coding line becoming the workhorse behind Microsoft's own products, not just a Copilot option.
Microsoft's internal results are modest in most apps. MAI Work Flash is "on par" with GPT-6 Luna in Excel and PowerPoint tasks and "similar to GPT 6.1 sol" in Word, using about 42% fewer input tokens per call. In Teams it beats GPT-6 Luna at tracking decisions and following custom instructions, with up to 28% fewer tokens. The clearest win is Power BI. On natural-language-to-DAX queries, MAI Work Flash scores 0.77 against 0.54 for GPT-6 Astra and 0.30 for Claude Opus 5.5.
The Rest of Microsoft's Ignite 2026 MAI Lineup
Microsoft announced what it calls "a family of seven new world class models" at Ignite. Besides MAI Code 2 and MAI Code 2 Flash, the list covers security, images, voice, transcription and routing.
| Model | What it does | Microsoft's claim | Availability |
|---|---|---|---|
| MAI Cyber 2 Flash | Vulnerability discovery and fixes | 92.3% on CyberGym-E2E vs 91.5% for GPT-6 Astra, at under 2% of the cost per task | Microsoft security products only, incl. MDASH from mid-December |
| MAI Image 2.7 | Image generation and editing | #1 on Artificial Analysis image editing, about $35 per 1,000 images (Flash about $10) | Public preview in Microsoft Foundry |
| MAI Realtime 1 | Live speech-to-speech | Beats GPT-Realtime 2.1, from $0.01 per audio minute | Foundry limited-access preview |
| Dragon Copilot Transcribe | Medical transcription on MAI Transcribe 2 | 10 million patient encounters a month | Live in Dragon Copilot |
| Decision 1.1 | Routes requests to the right model | 50x faster and 79x cheaper than GPT-6.1 Sol | Used by 4,000 customers, per Microsoft |
MAI Cyber 2 Flash and MAI Image 2.7
MAI Cyber 2 Flash comes with a disclosure worth reading. Microsoft found that 92 of CyberGym-E2E's 905 tasks overlapped with its training data. It dropped those tasks for every model and scored the rest. Microsoft also says it has not yet compared against GPT-6 Cyber, which OpenAI offers only to select customers.
MAI Image 2.7 is the one model in the lineup with an outside ranking behind it on launch day. Artificial Analysis, which runs the image editing leaderboard, posted its own result:
Microsoft AI has launched MAI-Image-2.7, the new #1 in AA-Image-Editing v2.1 at one sixth the price of the four models behind it
— Artificial Analysis (@ArtificialAnlys) November 17, 2026
MAI-Image-2.7 and MAI-Image-2.7-Flash both take new places on the Quality vs Price frontier in Image Editing and Text to Image, at $34.82 and $10.24 per 1k images.
Both perform best in Image Editing: MAI-Image-2.7 ranks #1 for Text or Symbol, Scene & Style and Enhancement & Restoration edits, and MAI-Image-2.7-Flash ranks #3 for Scene & Style edits.
Congratulations to @MicrosoftAI on the launch of MAI-Image-2.7 and MAI-Image-2.7-Flash!
See below for our detailed analysis of both models in Image Editing and Text to Image 🧵
MAI Licensed Weights
MAI Licensed Weights lets a company customize a MAI model with its own data and run it inside its own Azure environment. Microsoft says the customer's data and adaptations stay theirs. It is "coming soon" through Microsoft's Frontier Company engagements, not as a self-serve product.
The customer results Microsoft shared in its Licensed Weights announcement all run on the new coding models. AT&T customized MAI Code 2 Flash and MAI Code 2 for 5G network analytics and, on standard tasks, beat GPT-6.1 Sol at one fiftieth of the cost. Woven by Toyota tuned MAI Code 2 Flash on coding-standard fixes and got ready-to-merge fixes in 44% of the cases where its existing frontier models needed a human. Alaska Airlines raised maintenance-report classification accuracy from 81% to 91%.
Why Microsoft Is Building Its Own Coding Models
Microsoft is building its own models so it can sell cheaper ones next to OpenAI's and Anthropic's, not instead of them. GitHub Copilot already offers GPT-6.2, Claude Opus 5.5 and Gemini 4 models in the same picker. MAI Code 2 gives Microsoft a model of its own in that list, priced well below the frontier.
Mustafa Suleyman framed it as a system rather than a single release. "This isn't just one model. It's not a flash in the pan," he told Reuters. "This is a system, a hill-climbing machine that we have built, which co-optimizes performance and price."
The keynote spelled out the trade. A model built to push science forward, Microsoft argued, makes you "pay for that capability on every task, even for a routine bug fix." MAI Code 2 is the opposite bet: give up the last few points on hard problems and win on the bill.
Should You Use MAI Code 2?
Use MAI Code 2 for routine Copilot work while the promo lasts, and keep a frontier model for the hard tasks. On Microsoft's four-benchmark average, the gap to GPT-6.1 Sol is under one point. Until 17 May 2027, the cost per task is less than half.
Do not swap it in for the hard jobs. Microsoft itself says MAI Code 2 is not tuned for very long agentic runs. Its SWE Atlas score trails Claude Opus 5.5 by 14 points. A long refactor, or an agent left running for an hour, still belongs on Opus or a GPT-6.2 model. Anyone who picks MAI-Code-1.1-Flash by hand should move to MAI Code 2 Flash before the 18 December cutoff. Outside the editor, you face the same choice between labs in every chat app. Fello AI lets you ask GPT, Claude and Gemini the same question on your Mac without a second subscription.
The verdict: MAI Code 2 is a real price cut for everyday coding, but the 40% headline is a six-month offer. Microsoft's own numbers say the lasting deal is against Claude Opus 5.5, at about a tenth of the cost per task. Against GPT-6.1 Sol, it settles at about a fifth cheaper. If you want a free starting point instead, see our list of the best free AI for coding.
Frequently Asked Questions
Is MAI Code 2 free?
No. MAI Code 2 needs a paid GitHub Copilot plan: Pro, Pro+, Max, Business or Enterprise. The smaller MAI Code 2 Flash is also available on Copilot Free and Copilot Student.
How much does MAI Code 2 cost?
MAI Code 2 costs $0.50 per million input tokens and $2.50 per million output tokens in GitHub Copilot until 17 May 2027. After that it costs $1 and $5. MAI Code 2 Flash costs $0.05 and $0.25 until the same date, then $0.10 and $0.50.
Is MAI Code 2 better than Claude Opus 5.5?
No, but it is much cheaper. In Microsoft's tests, MAI Code 2 averages 63.2% across four coding benchmarks against 65.3% for Claude Opus 5.5. It costs $0.309 per task at the promo price against $6.306. Opus 5.5 leads clearly on SWE Atlas.
What happens to MAI-Code-1.1-Flash?
GitHub retires MAI-Code-1.1-Flash across all Copilot experiences on 18 December 2026. The suggested replacement is the latest MAI Code Flash model, which is now MAI Code 2 Flash.
Can I use MAI Code 2 outside GitHub Copilot?
Not yet for MAI Code 2 itself. MAI Code 2 Flash is in a limited-access preview in Microsoft Foundry. Microsoft says releases on OpenRouter, Vercel, Hermes, OpenCode and Kilo are coming soon. Enterprises will also be able to license and customize MAI weights through Microsoft's MAI Licensed Weights program.
The featured image for this article was created with the help of AI.