MAI Code 2 is Microsoft's new coding model, and in GitHub Copilot it costs $0.50 per million input tokens and $2.50 per million output tokens until May 17, 2027, then $1.00 and $5.00. By our arithmetic, that is one eighth of Claude Opus 5.5's per-token price ($4 and $20) and one quarter of GPT-6.1 Sol's ($2 and $10) at the promotional rate. Microsoft says MAI Code 2 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. Those are vendor-reported numbers, and the cost side uses the 50% launch price.
Microsoft announced MAI Code 2 and MAI Code 2 Flash at Ignite on November 17, 2026. Below you get the prices from the official pricing pages (all checked on November 20, 2026), the per-benchmark scores behind Microsoft's averages, what the cost claim turns into once the promotion ends, and where the model fits if your team already runs Claude Code.
What did Microsoft launch?
Two models, both built for everyday coding work, announced in a Microsoft AI post titled "MAI Code 2 and Code 2 Flash: Everyday coding at scalable prices" (Microsoft AI, November 17, 2026):
- MAI Code 2, the larger model, pitched against GPT-6.1 Sol and Claude Opus 5.5.
- MAI Code 2 Flash, the smaller one for high-volume work, pitched against GPT-6 Luna and Claude Haiku 5.5.
Where you can use them today:
| Where | MAI Code 2 | MAI Code 2 Flash | Source |
|---|---|---|---|
| GitHub Copilot | Available on Pro, Pro+, Max, Business, Enterprise | Available on Free, Student, Pro, Pro+, Max, Business, Enterprise | GitHub changelog, Nov 17, 2026 |
| Microsoft Foundry | Not stated | Limited-access preview | Microsoft AI |
| OpenRouter, Vercel, Hermes, OpenCode, Kilo | "Soon", no date | "Soon", no date | Microsoft AI |
In Copilot, the changelog lists the model picker in VS Code, Visual Studio, Copilot CLI, the cloud agent, the Copilot app, github.com, GitHub Mobile, JetBrains, Xcode and Eclipse. Both models are "billed under usage-based billing," and the changelog adds that they are available to eligible subscribers on annual plans. GitHub's pricing page lists both as GA.
Watch out: On the Free plan you get MAI Code 2 Flash only. Both the changelog and the Copilot plans page mark MAI Code 2 as not included on Free, so it starts at Copilot Pro.
One thing Microsoft says plainly: the models are "explicitly not yet tuned for harder problems, like fundamental science research and very long-horizon agentic work" (Microsoft AI). Keep that sentence in mind for the Claude Code section below.
How much does MAI Code 2 cost?
All prices per million tokens, from GitHub's Copilot models and pricing page, checked November 20, 2026. We picked that page because it lists every model on the same footing.
| Model | Input | Cached input | Output | Note |
|---|---|---|---|---|
| MAI Code 2 (promo) | $0.50 | $0.025 | $2.50 | Until May 17, 2027 |
| MAI Code 2 (from May 17, 2027) | $1.00 | $0.05 | $5.00 | Standard price |
| MAI Code 2 Flash (promo) | $0.05 | $0.005 | $0.25 | Until May 17, 2027 |
| MAI Code 2 Flash (from May 17, 2027) | $0.10 | $0.01 | $0.50 | Standard price |
| Claude Opus 5.5 | $4.00 | $0.20 | $20.00 | Plus $5.00 cache write |
| Claude Sonnet 5.5 | $2.00 | $0.10 | $10.00 | Plus $2.50 cache write |
| Claude Haiku 5.5 (prompt up to 100K) | $0.10 | $0.01 | $0.50 | $0.50 / $2.50 over 100K |
| GPT-6.1 Sol (up to 272K) | $2.00 | $0.10 | $10.00 | $4.00 / $15.00 over 272K; plus $2.50 cache write |
| GPT-6 Luna (up to 272K) | $0.10 | $0.01 | $0.50 | $0.20 / $0.75 over 272K; plus $0.125 cache write |
The Claude and GPT rows match what the model makers charge on their own APIs. Anthropic lists Opus 5.5 at $4 input and $20 output, Sonnet 5.5 at $2 and $10, and Haiku 5.5 at $0.10 and $0.50 for prompts up to 100,000 tokens (Anthropic pricing, checked November 20, 2026). OpenAI lists GPT-6.1 Sol at $2.00 input and $10.00 output, and GPT-6 Luna at $0.10 and $0.50 (OpenAI API pricing, checked November 20, 2026). The over-272K rates in the GPT rows are GitHub's; OpenAI's API pricing page lists no long-context tier for these two models.
Copilot converts usage into AI credits, where one credit is $0.01. Individual plans come with a monthly credit allowance, and Business and Enterprise plans pool per-user allowances across the billing entity.
At the time of writing, MAI Code 2 has no published per-token price outside GitHub. Microsoft's announcement gives none, including for Flash's Foundry preview.
What do the benchmarks actually say?
Microsoft's headline is an average over four benchmarks: DeepSWE 1.1, Terminal-Bench 2.1, SWE-Atlas and FrontierCode v1.1. The per-benchmark scores below come from the charts in Microsoft's announcement, and InfoWorld reported the same figures on November 18, 2026. All are vendor-reported: no independent lab has published a reproduction yet.
| Benchmark (vendor-reported) | MAI Code 2 | GPT-6.1 Sol | Claude Opus 5.5 |
|---|---|---|---|
| DeepSWE 1.1 | 73.2% | 74.6% | 69% |
| Terminal-Bench 2.1 | 84.6% | 85% | 78.7% |
| SWE-Atlas | 43.1% | 47.4% | 57.5% |
| FrontierCode v1.1 | 52% | 49.4% | 56% |
| Average | 63.2% | 64.1% | 65.3% |
The ratios hold up by our arithmetic: 63.2 divided by 64.1 is 98.6%, and 63.2 divided by 65.3 is 96.8%. But the average hides two different stories. The point gaps below are our subtraction from Microsoft's table.
- Against GPT-6.1 Sol, MAI Code 2 is close on every test. It trails by less than 1.5 points on DeepSWE and Terminal-Bench, by 4.3 on SWE-Atlas, and leads by 2.6 on FrontierCode.
- Against Claude Opus 5.5, the gap depends on the test. MAI Code 2 leads by 5.9 points on Terminal-Bench 2.1 and 4.2 on DeepSWE 1.1. Opus leads by 14.4 points on SWE-Atlas and 4 on FrontierCode. A 96% average comes from a big loss on one test and a smaller loss on another, partly offset by wins on the other two.
How the tests were run matters too. Microsoft's footnotes say every model ran at maximum reasoning effort. DeepSWE, Terminal-Bench and SWE-Atlas ran on Microsoft's MSBench harness through the GitHub Copilot CLI, without subagents. FrontierCode scores and costs came from Artificial Analysis. So on three of the four benchmarks, Opus 5.5 was measured inside the Copilot CLI with subagents off, not inside Claude Code. The FrontierCode figures come from Artificial Analysis's own runs. Microsoft's own caveat: "Performance and cost comparisons vary by workload, model configuration, reasoning effort, and pricing assumptions."
What we take from it: if your work looks like Terminal-Bench (shell tasks, scripts, environment setup), the reported numbers favour MAI Code 2 over Opus 5.5. If it looks like SWE-Atlas, Opus leads by a wide margin. Either way, a vendor table is a reason to run your own test, not a reason to switch.
Is it really 5% of the cost of Claude Opus 5.5?
Only at the launch price, and only on Microsoft's test set. Microsoft's footnote says "MAI Code 2 cost is shown at its 50% promotional price." Microsoft's chart gives the average cost per task behind the claim: $0.309 for MAI Code 2, $0.761 for GPT-6.1 Sol and $6.306 for Claude Opus 5.5.
The table below doubles MAI Code 2's cost to show the standard price that starts May 17, 2027. That column is our arithmetic, not Microsoft's. It assumes each model uses the same number of tokens per task as in Microsoft's runs and that the other models keep today's prices.
| Comparison | Promo price (Microsoft's claim) | Standard price (our arithmetic) |
|---|---|---|
| MAI Code 2 vs GPT-6.1 Sol | $0.309 vs $0.761 = 41% | $0.618 vs $0.761 = 81% |
| MAI Code 2 vs Claude Opus 5.5 | $0.309 vs $6.306 = 4.9% | $0.618 vs $6.306 = 9.8% |
Two things follow from that table.
- Against Opus, the saving survives the promotion. At the standard price, MAI Code 2 would still cost about a tenth as much per task on these benchmarks. Per token it is a quarter of Opus's price at standard rates, so our inference is that the rest of the gap comes from Opus using more tokens per task at maximum effort in these runs.
- Against GPT-6.1 Sol, most of the saving is the discount. Per token, standard MAI Code 2 costs half as much as Sol. Its per-task cost would be 81% of Sol's, which suggests it used more tokens per task than Sol in these runs.
How does MAI Code 2 Flash compare with Claude Haiku 5.5 and GPT-6 Luna?
Microsoft says Flash gives "comparable performance to GPT-6 Luna and Claude Haiku 5.5 at 65% and 20% of the cost," measured on three benchmarks (DeepSWE 1.1, Terminal-Bench 2.1 and SWE-Atlas) at maximum reasoning effort, with Flash's cost again at its 50% promotional price. The scores, from Microsoft's chart:
| Benchmark (vendor-reported) | MAI Code 2 Flash | GPT-6 Luna | Claude Haiku 5.5 |
|---|---|---|---|
| DeepSWE 1.1 | 59.9% | 59.6% | 46% |
| Terminal-Bench 2.1 | 78.3% | 69.3% | 73.4% |
| SWE-Atlas | 44.8% | 36.4% | 42.1% |
| Average | 61% | 55.1% | 53.9% |
| Avg cost per task | $0.052 (promo) | $0.080 | $0.319 |
Here Flash leads on every test, which is a stronger result than the big model's. The cost picture changes once the promotion ends:
- Against GPT-6 Luna, $0.052 vs $0.080 is 65% at the promo price. Doubled to the standard price, $0.104 would be about 130% of Luna's cost by our arithmetic, so Flash would cost more per task than Luna, while still scoring higher in Microsoft's runs.
- Against Claude Haiku 5.5, Microsoft's own chart puts Flash at $0.052 against $0.319 per task, which is 16% by our arithmetic, not the 20% in its text, and Microsoft doesn't explain the gap. Doubled to the standard price, $0.104 would be about 33% of Haiku's cost (our arithmetic).
Per token, Flash at standard price ($0.10 / $0.50) costs exactly what Haiku 5.5 and GPT-6 Luna cost at short context. So the per-task gap comes mostly from how many tokens each model spends, plus the higher long-prompt rates (Haiku 5.5 over 100K, Luna over 272K) and cache-write charges that Flash's price list doesn't include.
Where does MAI Code 2 fit if you use Claude Code?
Claude Code is Anthropic's coding agent: a harness with its own tools, subagents, hooks, skills and CLAUDE.md memory, running Claude models. MAI Code 2 is a model. Today you use it through GitHub Copilot's harness. So for most teams the real choice is Claude Code with Claude models against GitHub Copilot with a model picker that now includes MAI Code 2. We compared the two tools in more depth in GitHub Copilot vs Claude Code.
Subscription prices, checked November 20, 2026:
| Plan | Price | What it includes | Source |
|---|---|---|---|
| Claude Pro | $20/month ($17/month billed annually) | Claude Code included | claude.com/pricing |
| Claude Max | $100/month | Claude Code, 5x Pro usage | claude.com/pricing |
| Claude Ultra | $200/month | Claude Code, 10x Pro usage weekly, no 5-hour limit | claude.com/pricing |
| Claude Team, standard / premium seat | $25 / $125 per seat monthly ($20 / $100 annually) | Claude Code included | claude.com/pricing |
| Copilot Pro | $10/month | $15 monthly credits | GitHub Copilot plans |
| Copilot Pro+ | $39/month | $70 monthly credits | GitHub Copilot plans |
| Copilot Max | $100/month | $200 monthly credits | GitHub Copilot plans |
Our read on where each one fits, based on what Microsoft itself says the model is and isn't for:
MAI Code 2 in Copilot is worth a trial for
- High-volume everyday tickets: small fixes, tests, scripts, refactors inside one module.
- Shell and environment work, where the reported Terminal-Bench numbers are strongest.
- Teams already paying for Copilot who want a cheaper default model while the promotion runs.
Claude Code with Opus 5.5 still makes sense for
- Long agent runs across a whole repo. Microsoft says MAI Code 2 is "not yet tuned" for very long-horizon agentic work.
- Work that leans on Claude Code's harness: subagents, hooks, skills, project memory. Microsoft's Copilot CLI runs turned subagents off.
- Tasks that look like SWE-Atlas, where Opus led by 14.4 points in Microsoft's own table.
Running both is reasonable. Some teams use Copilot in the editor and Claude Code in the terminal. If rate limits are why you're looking at cheaper models, our post on Claude Code rate limits covers what changed and how to work within them. If you want help setting up Claude Code properly before deciding, that is what our Claude Code development team does.
How to test MAI Code 2 on your own code
Vendor benchmarks are a starting point. A one-week test on your own repo tells you more.
- Pick 10 to 20 closed tickets from the last quarter that have a clear pass check (tests, a reviewed diff).
- Run each one with MAI Code 2 in Copilot and with your current setup, at the same reasoning effort.
- Record pass or fail, reviewer edits needed, and the credits or tokens each run used.
- Price the MAI Code 2 runs at both the promo and the standard rate, since the standard rate applies from May 17, 2027.
- Decide per task type, not per model. You may end up with a cheaper model for tests and scripts and a stronger one for cross-repo changes.
If you don't have the people to run that test, our AI developers can run this kind of model and tool evaluation for your team.
FAQ
How much does MAI Code 2 cost?
In GitHub Copilot, MAI Code 2 costs $0.50 per million input tokens, $0.025 cached and $2.50 output until May 17, 2027. From then on it is $1.00, $0.05 and $5.00. MAI Code 2 Flash costs $0.05, $0.005 and $0.25 during the promotion, then $0.10, $0.01 and $0.50. Prices are from GitHub's Copilot pricing page, checked November 20, 2026.
Is MAI Code 2 better than Claude Opus 5.5?
Not on average, by Microsoft's own numbers. MAI Code 2 averaged 63.2% across four benchmarks against Opus 5.5's 65.3%. It led on Terminal-Bench 2.1 and DeepSWE 1.1, and trailed by 14.4 points on SWE-Atlas. All figures are vendor-reported.
Is MAI Code 2 really 5% of the cost of Opus?
On Microsoft's benchmark runs and at the 50% launch price, yes: $0.309 per task against $6.306. At the standard price from May 17, 2027, the same runs would put it near 10%, by our arithmetic. Your cost depends on how many tokens your tasks use.
Can I use MAI Code 2 in Claude Code?
Not as a supported option today. MAI Code 2 is available in GitHub Copilot, and Microsoft says OpenRouter, Vercel, OpenCode, Kilo and Hermes are coming "soon" with no date. Claude Code runs Claude models. If your team is moving to Claude Code, our Claude Code training covers the setup.
Which GitHub Copilot plans include MAI Code 2?
GitHub's changelog lists MAI Code 2 on Copilot Pro, Pro+, Max, Business and Enterprise, and MAI Code 2 Flash on every plan including Free and Student. The Copilot plans page agrees: on Free, MAI Code 2 is marked not included and Flash is included.
What is MAI Code 2 not good at?
Microsoft says the models are "explicitly not yet tuned for harder problems, like fundamental science research and very long-horizon agentic work." For long autonomous runs across a large codebase, a model built for that, such as Claude Opus 5.5 in Claude Code, is the safer pick until independent tests say otherwise.
Is MAI Code 2 available in Microsoft Foundry?
Only MAI Code 2 Flash, and only in limited-access preview, according to Microsoft's announcement. The announcement gives no Foundry price for it, and it says nothing about MAI Code 2 itself in Foundry.
Sources: Microsoft AI, "MAI Code 2 and Code 2 Flash: Everyday coding at scalable prices," Nov 17, 2026, GitHub changelog, Nov 17, 2026, GitHub Copilot models and pricing, GitHub Copilot plans, InfoWorld, Nov 18, 2026, Anthropic API pricing, Claude plans, OpenAI API pricing. Prices checked November 20, 2026.
Book a Free Strategy Call
Building this in production?
Walid runs a 30-min call to map your AI engineering team. Free, no slides.

Ex-IBM AI engineer and enterprise architect. Adel owns the technical architecture behind every automation and AI agent system AY Automate ships.
