MAI Code 2 and MAI Code 2 Flash at a glance
| MAI Code 2 | MAI Code 2 Flash | |
|---|---|---|
| Released | November 17, 2026 (Ignite), GA in GitHub Copilot | November 17, 2026 (Ignite), GA in GitHub Copilot |
| Context window | 1,050,000 tokens (Microsoft calls it 1M) | 1,050,000 tokens (Microsoft calls it 1M) |
| Max input / max output | 922,000 / 128,000 tokens | 922,000 / 128,000 tokens |
| Inputs | Text and images | Text and images |
| Reasoning effort | low, medium, high, xhigh, max | low, medium, high, xhigh, max |
| Knowledge cutoff | August 2026 | August 2026 |
| Copilot plans | Pro, Pro+, Max, Business, Enterprise | Free and Student (through auto model selection only), Pro, Pro+, Max, Business, Enterprise |
| Copilot price, promotional (until May 17, 2027) | $0.50 input / $2.50 output per 1M tokens | $0.05 input / $0.25 output per 1M tokens |
| Copilot price, from May 17, 2027 | $1.00 input / $5.00 output per 1M tokens | $0.10 input / $0.50 output per 1M tokens |
| Premium request multiplier (legacy annual plans only) | 1 | 0.33 |
| Outside Copilot | Not announced | Limited-access preview in Microsoft Foundry |
What MAI Code 2 is, and how it differs from MAI-Code-1.1-Flash
MAI Code 2 and MAI Code 2 Flash are the "MAI Code 2 family". Per Microsoft's model card, both share the same architecture family, training recipe and a similar data mix. The main difference is size. Both are decoder-only transformers with sparse Mixture-of-Experts layers and a built-in vision encoder, built on MAI-Base-2.
- MAI Code 2 is for complex work: deep debugging, multi-file changes, refactoring, repository-wide reasoning, UI and taking code to production. Microsoft says it is "at the frontier of price-performance for common enterprise coding tasks."
- MAI Code 2 Flash is tuned for low latency and low serving cost: code explanations, targeted edits, routine fixes and fast iteration. The card says it is less capable on complex, long-horizon tasks and may need more guidance.
- The predecessors were Flash-only. MAI-Code-1-Flash came out on June 2, 2026 at Build and was retired on September 10, 2026. MAI-Code-1.1-Flash followed on August 11, 2026, with 5B active parameters per the MAI Code 2 card.
- GitHub will retire MAI-Code-1.1-Flash across all Copilot experiences on December 18, 2026. GitHub's suggested replacement is "Latest MAI Code Flash model". MAI-Code-1.1-Flash is still listed as a preview in the Foundry catalog.
- Parameter counts: the model card gives two different sets of numbers for each model, so we leave them out.
- GitHub calls MAI models "continuously improving": behavior may change as new checkpoints ship.
MAI Code 2 benchmarks (Microsoft's own claims)
These are Microsoft's own numbers, not independent tests. Microsoft ran them on its internal MSBench platform with GitHub Copilot CLI v1.0.102, subagents off, max reasoning effort, averaged over 3 runs (single-attempt pass rate).
The headline comparisons average four benchmarks: DeepSWE 1.1, Terminal Bench 2.1, SWE-Atlas and FrontierCode v1.1. MAI costs are counted at the 50% promotional price. FrontierCode scores and costs come from Artificial Analysis. Microsoft says results vary by workload, configuration, reasoning effort and pricing assumptions.
| Claim | MAI Code 2 | MAI Code 2 Flash |
|---|---|---|
| vs GPT-6.1 Sol (4-benchmark average) | 98% of its performance at 40% of its cost | - |
| vs Claude Opus 5.5 (4-benchmark average) | 96% of its performance at 5% of its cost | - |
| vs GPT-6 Luna and Claude Haiku 5.5 (3 benchmarks) | - | "Comparable performance"; cost about 65% of GPT-6 Luna and about 20% of Haiku 5.5 |
| Blind frontend preference vs GPT-6 Sol (258 reviewers) | Preferred 64% of the time | - |
| DeepSWE 1.1 (model card, main table) | 73.2% | 59.9% |
| TerminalBench 2.1 (model card) | 84.6% | 78.3% |
| SWE Atlas (model card, main table) | 43.1% | 44.8% |
| SWE-bench (Verified, Pro or Multilingual) | Not published | Not published |
What the benchmark numbers don't tell you
- The model card prints other DeepSWE 1.1 (60.2% and 27.4%) and SWE Atlas (80.9% and 65.2%) figures in its safety section, without explaining the different setup. The table above uses the main benchmark table.
- On the main table Flash scores slightly above MAI Code 2 on SWE Atlas.
- Microsoft published no FrontierCode v1.1 score per model, and no per-benchmark numbers for Flash against its rivals, in the sources we checked.
- No SWE-bench score, and no speed figures: no tokens per second or latency numbers for either model.
- Each vendor runs its own benchmarks its own way. Don't line these numbers up against another vendor's tables as if they were run the same way.
- Microsoft compared against GPT-6.1 Sol. For OpenAI's newer models, see GPT-6.2 Astra and Sol.
MAI Code 2 pricing in GitHub Copilot
Copilot bills these models per token, converted to GitHub AI Credits (1 credit = $0.01). Prices are per 1M tokens, from GitHub's docs. The promotional prices run until May 17, 2027. Choosing the 1M context or a higher reasoning level uses more credits.
| Model | Input | Cached input | Output |
|---|---|---|---|
| MAI Code 2, promotional | $0.50 | $0.025 | $2.50 |
| MAI Code 2, from May 17, 2027 | $1.00 | $0.05 | $5.00 |
| MAI Code 2 Flash, promotional | $0.05 | $0.005 | $0.25 |
| MAI Code 2 Flash, from May 17, 2027 | $0.10 | $0.01 | $0.50 |
MAI Code 2 vs Claude Opus 5.5
Microsoft's headline claim is that MAI Code 2 reaches 96% of Claude Opus 5.5's performance at 5% of its cost. Read it with the caveats Microsoft gives.
- It is an average over four benchmarks Microsoft picked, run on Microsoft's own harness with Copilot CLI at max reasoning.
- The 5% cost figure uses MAI Code 2's promotional price, which doubles on May 17, 2027.
- We couldn't find a per-benchmark breakdown for Opus 5.5 in Microsoft's sources, so the 96% can't be checked line by line.
- Where they run: MAI Code 2 is in GitHub Copilot only, with no API announced. Opus 5.5 is Anthropic's model, used in Claude Code and on the Claude API, where it lists at $4 per million input tokens and $20 per million output.
- Anthropic recommends Opus 5.5 as the starting point for most work and its new Claude Fable 5.2 for the hardest jobs.
- Microsoft says MAI Code 2 is not yet tuned for very long-horizon agentic work. If your tasks run for hours, test on your own repo before you switch.
How to use MAI Code 2 in GitHub Copilot
Update your client
Visual Studio needs 17.14.6 or later. GitHub lists the minimum for VS Code, JetBrains, Xcode and Eclipse as "TBD", so update to the latest Copilot extension or plugin.
Pick the model
Open Copilot Chat or agent mode in VS Code, or Copilot CLI, and choose MAI Code 2 or MAI Code 2 Flash in the model picker. The models are also in Copilot cloud agent, the GitHub Copilot app, github.com, GitHub Mobile, Visual Studio, JetBrains, Xcode and Eclipse.
Turn on 1M context or set reasoning, if you need it
The 1M-token context is available only in VS Code and Copilot CLI. Configurable reasoning works in VS Code, Copilot CLI and Copilot cloud agent. Both cost more credits. GitHub doesn't state the default context size for MAI models.
On Business or Enterprise, check the model policy
If "Default availability for released models" is on, the models are enabled unless an admin disabled them. To enable them yourself as an organization owner: profile picture → Organizations → your org → Settings → Code, planning, and automation → Copilot → Models, then set MAI Code 2 and MAI Code 2 Flash to Enabled. A shield icon means your enterprise owner controls the setting.
On Free or Student, use Auto
These plans get MAI Code 2 Flash only through auto model selection. You can't pick it by name, and MAI Code 2 isn't offered.
Source: GitHub changelog, MAI Code 2 and MAI Code 2 Flash in GitHub Copilot
Can you use MAI Code 2 outside Copilot?
- MAI Code 2: not yet. Microsoft names GitHub Copilot as its only channel and has announced no API or Foundry offering.
- MAI Code 2 Flash: in a limited-access preview in Microsoft Foundry. You create a deployment and call it through the OpenAI-compatible Responses API with Azure authentication. Microsoft hasn't published how to request access, a Foundry price or a model ID in the sources we checked.
- Microsoft says it will "soon" release on Open Router, Vercel, Hermes, Open Code and Kilo. There is no date, and the sentence may refer to Flash only.
- Weights are not released. The models run only on Azure.
- Nothing official says you can use either model in Codex, Claude Code or Cursor.
Limits Microsoft states
- Not yet tuned for fundamental science research or very long-horizon agentic work.
- Not meant for consumer chat, or for medical, legal or financial advice.
- In long sessions the models may repeat tool calls, drift, or stop before verifying their work. Quality may drop near the max context.
- Optimized for English code production. Other languages are not extensively evaluated.
- Microsoft rates their alignment as similar to GPT-5.6 Sol and below GPT-6.
- Under Microsoft's safety framework, CBRN and offensive cyber risk are rated High, with classifiers, logging and abuse monitoring as mitigations.
Want an agent that works a backlog on a schedule?
MAI Code 2 lives inside GitHub Copilot for now. Crewmark turns coding-agent sessions into AI employees, each with a role, a goal, working hours and a heartbeat, taking issues from a board. Claude Code and Codex employees run natively; other agents, such as OpenCode and DeepSeek Harness, connect through ACP (acpx), and GitHub Copilot's CLI is among the agents acpx can drive. Copilot users can also hand work to the Copilot agent from chat: see GitHub Copilot in Slack.
Pick a role
Such as the AI software engineer for a backlog of bugs and small features.
Choose the agent
Run the employee on Claude Code, Codex or another agent; see Claude Code vs Codex if you're unsure which.
Let it run
It wakes on its heartbeat and takes the next issue from the board; see how to run Claude Code 24/7.
FAQ
What is MAI Code 2?
A coding model from Microsoft AI, announced at Ignite on November 17, 2026 and trained from scratch by Microsoft. It's built for complex work such as deep debugging, multi-file changes and repository-wide reasoning. It's available in GitHub Copilot on Pro, Pro+, Max, Business and Enterprise.
What is MAI Code 2 Flash?
The smaller, cheaper model in the same family, tuned for low latency: explanations, targeted edits and routine fixes. In Copilot it costs $0.05 input and $0.25 output per million tokens until May 17, 2027. It's also in a limited-access preview in Microsoft Foundry.
What are MAI Code 2's benchmarks?
Microsoft's own figures: 98% of GPT-6.1 Sol's performance at 40% of its cost, and 96% of Claude Opus 5.5's at 5% of its cost, averaged over DeepSWE 1.1, Terminal Bench 2.1, SWE-Atlas and FrontierCode v1.1 at the promotional price. The model card's main table gives 73.2% on DeepSWE 1.1 and 84.6% on TerminalBench 2.1. Microsoft published no SWE-bench score.
Is MAI Code 2 better than Claude Opus 5.5?
Microsoft doesn't claim that. It claims 96% of Opus 5.5's performance at 5% of the cost, on four benchmarks it chose, using its promotional price. Opus 5.5 runs in Claude Code and on Anthropic's API; MAI Code 2 runs only in GitHub Copilot for now.
What happens to MAI-Code-1.1-Flash?
GitHub will retire it across all Copilot experiences on December 18, 2026, and suggests the latest MAI Code Flash model instead. It came out on August 11, 2026; the earlier MAI-Code-1-Flash was retired on September 10, 2026.
Can I use MAI Code 2 through an API or in other coding agents?
Not MAI Code 2: no API has been announced. MAI Code 2 Flash is in a limited Foundry preview through the Responses API. Microsoft says it will "soon" come to Open Router, Vercel, Hermes, Open Code and Kilo, without a date.
What is the context window of MAI Code 2?
1,050,000 tokens per the model card, with up to 922,000 input and 128,000 output tokens. In Copilot, the 1M window is available only in VS Code and Copilot CLI.
What are people on Reddit saying about MAI Code 2?
This page sticks to what Microsoft and GitHub have published, not forum threads. Microsoft hasn't published speed figures or a SWE-bench score, so testing it on your own repo in Copilot is the most reliable check.
Checked against Microsoft AI's announcement, the MAI Code 2 model card, GitHub's changelog and the GitHub Copilot docs on November 18, 2026.