MAI Code 2 and MAI Code 2 Flash: one family, two sizes
Microsoft's two new coding models, side by side: official specs, the benchmark tables it published, per-token prices, and how to select either one in GitHub Copilot.
MAI Code 2 and Code 2 Flash in short
- Microsoft announced both models on 17 November 2026. They shipped the same day in GitHub Copilot as generally available models, and Flash is also in limited-access preview in Microsoft Foundry.
- They are one family: same architecture, same training recipe, same data mix — the difference is size. Code 2 has 77B active parameters, Flash has 12B, and both take a 1,050,000-token context.
- On Microsoft's own benchmarks Code 2 averages 63.2% pass at $0.309 per task, Flash 61.0% at $0.052. Microsoft puts Code 2 at "98% of the performance of GPT-6.1 Sol at 40% of the cost".
- Copilot access differs: Flash is on Free and Student too, Code 2 starts at Pro. Both bill as usage-based AI credits — Flash at $0.05 in / $0.25 out per 1M tokens during the promotional period.
Introducing MAI Code 2 and MAI Code 2 Flash
Sourcemicrosoft.ai2026-11-20
MAI Code 2 and MAI Code 2 Flash — Model Card (PDF)
Sourcemicrosoft.ai2026-11-20
MAI Code 2 and MAI Code 2 Flash in GitHub Copilot
Sourcegithub.blog2026-11-20
Models and pricing for GitHub Copilot
Sourcedocs.github.com2026-11-20
Supported AI models in GitHub Copilot
Sourcedocs.github.com2026-11-20
Specs, benchmarks, prices and plan eligibility on this page were read from the sources above on 2026-11-20. Where a number is not on those pages, it is not on this page either.
What MAI Code 2 and MAI Code 2 Flash are
They are Microsoft AI's agentic coding models, announced on 17 November 2026. Microsoft introduces them as "two new coding models built for everyday enterprise coding tasks, delivering leading performance at a fraction of the cost of competing models".
The split is deliberate. Code 2 is the larger model, built for "taking code to production, deep debugging, multi-file implementation, repo-wide reasoning, and UI". Flash is "designed for high-volume work such as code explanations, targeted edits, routine fixes, and rapid iterations, while bringing quality at a much lower cost".
Both were pre-trained and post-trained by Microsoft from scratch, and both are hosted only: the model card states that no local hardware is required and that model weights are not distributed. At release they are "optimized for English code production only", and Microsoft says plainly that they are "explicitly not yet tuned for harder problems, like fundamental science research and very long-horizon agentic work".
Official specifications, side by side
From the model card (last updated 17 November 2026) and GitHub's supported-models reference, both read on 2026-11-20.
| Spec | MAI Code 2 | MAI Code 2 Flash |
|---|---|---|
| Architecture | Decoder-only transformer with sparse mixture-of-experts layers and an integrated vision encoder | Same architecture family and training recipe, smaller base |
| Parameters | 77B active · 3B embedded · ~2.8T total | 12B active · ~398B total |
| Context | 1,050,000 tokens (922,000 input / 128,000 output) | 1,050,000 tokens (922,000 input / 128,000 output) |
| Inputs | Text and images | Text and images |
| Reasoning effort | Configurable: low, medium, high, xhigh, max | Configurable: low, medium, high, xhigh, max |
| Knowledge cut-off | August 2026 | August 2026 |
| Release | 17 November 2026 — GA in GitHub Copilot | 17 November 2026 — GA in GitHub Copilot; limited-access preview in Microsoft Foundry |
| Copilot plans | Pro, Pro+, Max, Business, Enterprise | Free, Student, Pro, Pro+, Max, Business, Enterprise |
| Price per 1M tokens (promo) | $0.50 input · $0.025 cached · $2.50 output | $0.05 input · $0.005 cached · $0.25 output |
| Weights | Hosted service on Azure GPUs; weights not distributed | Hosted service on Azure GPUs; weights not distributed |
The model card summary table gives parameter counts as ranges (50B–100B with 80B active for Code 2, 5B–15B with 12B active for Flash), while section 1 gives the exact figures used above. Where the two disagree, the detailed section is quoted.
The benchmarks Microsoft published
Pass rate at max reasoning effort, on Microsoft's own evaluation harness. Flash was not run on FrontierCode, so that row is blank for it — the averages are not comparable across columns.
| Benchmark | MAI Code 2 | MAI Code 2 Flash |
|---|---|---|
| DeepSWE 1.1 | 73.2% | 59.9% |
| Terminal-Bench 2.1 | 84.6% | 78.3% |
| SWE Atlas | 43.1% | 44.8% |
| FrontierCode v1.1 | 52.0% | — |
| Average (the rows above) | 63.2% | 61.0% |
| Average cost per task | $0.309 | $0.052 |
Method, from the model card: every coding benchmark ran on MSBench with the GitHub Copilot CLI (v1.0.102) harness, subagents disabled, at max reasoning effort, averaged over three independent runs per task; infrastructure errors count as failures. Costs are Microsoft's figures at the 50% promotional price.
Averaged over DeepSWE 1.1, Terminal-Bench 2.1, SWE Atlas and FrontierCode v1.1, Microsoft puts MAI Code 2 at "98% of the performance of GPT-6.1 Sol at 40% of the cost, and 96% of Claude Opus 5.5's performance at just 5% of the cost".
For Flash, over DeepSWE 1.1, Terminal-Bench 2.1 and SWE Atlas, Microsoft claims "comparable performance to GPT-6 Luna and Claude Haiku 5.5 at 65% and 20% of the cost".
How the comparisons were drawn
The two tables below are the announcement's own charts. They use different benchmark sets, so compare rows inside a table, not across them.
MAI Code 2 against frontier models
Four benchmarks (DeepSWE 1.1, Terminal-Bench 2.1, SWE Atlas, FrontierCode v1.1): average pass rate and average cost per task in USD, all models at max reasoning effort.
| Model | Avg. pass rate | Avg. cost / task |
|---|---|---|
| MAI Code 2 | 63.2% | $0.309 |
| GPT-6.1 Sol | 64.1% | $0.761 |
| Claude Opus 5.5 | 65.3% | $6.306 |
MAI Code 2 Flash against lightweight models
Three benchmarks (DeepSWE 1.1, Terminal-Bench 2.1, SWE Atlas): average pass rate and average cost per task in USD.
| Model | Avg. pass rate | Avg. cost / task |
|---|---|---|
| MAI Code 2 Flash | 61.0% | $0.052 |
| GPT-6 Luna | 55.1% | $0.080 |
| Claude Haiku 5.5 | 53.9% | $0.319 |
What Copilot charges per token
Copilot prices model usage per token and converts the total into AI credits at 1 credit = $0.01 USD. These are the published rates per 1 million tokens.
| Model | Input | Cached input | Output |
|---|---|---|---|
| MAI Code 2 | $0.50 | $0.025 | $2.50 |
| MAI Code 2 Flash | $0.05 | $0.005 | $0.25 |
The promotional price ends on 17 May 2027
Both models carry a 50% promotional price until 17 May 2027. Starting 17 May 2027 the list price doubles: Code 2 at $1.00 input / $0.05 cached / $5.00 output, Flash at $0.10 / $0.01 / $0.50 per 1M tokens. Microsoft's benchmark cost figures are quoted at the promotional price.
What a request actually costs you
Every Copilot plan includes an AI-credit allowance, and usage past it is billed at the rates above. Copilot Free and Student include an allowance rather than unlimited Flash. Code completions and next-edit suggestions are not billed in AI credits; chat and agent requests are.
How to use them in GitHub Copilot
Four steps, all of them documented — no workarounds needed.
- 1
Open the model picker in any Copilot client
In VS Code, open Copilot Chat and use the model dropdown at the bottom of the chat view. The same picker exists in Visual Studio, the JetBrains IDEs, Xcode, Eclipse, Copilot CLI, the Copilot app, the cloud agent and on github.com.
- 2
Check that your plan includes the model
Flash is offered on Free, Student, Pro, Pro+, Max, Business and Enterprise. Code 2 starts at Pro. Both are available to eligible subscribers on annual plans as well.
- 3
If you are an admin, check the model policy
Business and Enterprise administrators manage MAI models through the model policy in Copilot settings. Under default model enablement, new models are enabled automatically unless an administrator turned off the global default or disabled the model itself.
- 4
Watch the AI-credit meter
These models are billed under usage-based billing, so a long agent session on Code 2 costs real money. Set a budget, keep routine work on Flash, and reserve Code 2 for the tasks that need the extra depth.
Which clients list them
The changelog says both models can be selected in VS Code, Visual Studio, Copilot CLI, the Copilot cloud agent, the Copilot app, github.com, GitHub Mobile on iOS and Android, the JetBrains IDEs, Xcode and Eclipse. The supported-models reference also flags both for the 1M-token context window (VS Code and Copilot CLI only) and configurable reasoning levels.
The other routes in, and what is actually confirmed
Copilot is the only generally available channel today. Everything else is either a limited preview or a promise.
Microsoft Foundry — Flash only, limited access
Flash is in limited-access preview in Microsoft Foundry. The model card describes the integration: create a MAI Code 2 Flash deployment and call it through the OpenAI-compatible Responses API, with standard Azure authentication, quotas and content filtering.
OpenRouter, Vercel, Hermes, OpenCode and Kilo — announced, not shipped
The announcement says exactly this: "Soon, we'll be releasing them on OpenRouter, Vercel, Hermes, OpenCode, and Kilo." No date, no pricing and no API model IDs have been published, so nothing beyond that sentence is verifiable today.
That is Microsoft stating an intent to release, not a partnership announcement: no third-party provider has confirmed availability, a model ID or a price. Until one does, GitHub Copilot and Microsoft Foundry are the only confirmed routes.
Which of the two to pick
The honest answer is "both, by task" — the family was designed that way, and the price gap is large enough to make routing worth the effort.
MAI Code 2 for production-shaped work
Deep debugging, multi-file implementation, refactors, repo-wide reasoning and UI. In blind human side-by-side evaluations across interfaces, games, data visualisations, 3D scenes and SVG graphics, developers preferred Code 2's output over GPT-6 Sol 64% of the time (258 annotators).
MAI Code 2 Flash for volume
Code explanations, targeted edits, routine fixes, test scaffolding and fast iteration. It averaged 61.0% pass against Code 2's 63.2% while costing about a sixth as much per benchmark task.
Do the arithmetic before you standardise
At the promotional rate Flash is ten times cheaper on input and output tokens. If your workload is mostly explain / rename / small-fix, the quality gap rarely pays for the price gap; if it is mostly agentic multi-file work, it usually does.
What Microsoft says the models cannot do
Quoted from the known-limitations section of the model card, not from third-party complaints.
- 1Knowledge stops at the pre-training cut-off in August 2026, so recent libraries, APIs and vulnerabilities may be unknown.
- 2In long agentic sessions the models can repeat tool calls, under-use the existing code base, drift from instructions or stop before fully verifying their work; quality degrades as context approaches its maximum length.
- 3"The smaller model, MAI Code 2 Flash is less capable than MAI Code 2 on complex, long-horizon tasks and may need more guidance."
- 4Higher reasoning effort raises latency and token spend and "does not always improve results".
- 5Both models are "optimized for English code production only" at release, and Microsoft advises against deploying generated code to production, safety-critical or regulated systems without human review and testing.
