MAI Code 2 is Microsoft's new coding model in GitHub Copilot, priced at a quarter of GPT-6.1 Sol per token, and it is probably already switched on for your developers. Its launch price is a 50% promotion that ends on May 17, 2027, and Microsoft's claim of "98% of the performance of GPT 6.1 Sol at 40% of the cost" was measured mostly on Microsoft's own harness, at that discounted price. If you run Copilot Business or Enterprise, decide this month whether MAI Code 2 is a default, an option or off, based on your own tasks rather than the launch chart, and budget for the price after May.
The model is good news for a Copilot budget. On GitHub's own price list, MAI Code 2 costs a quarter of GPT-6.1 Sol per token during the promotion and half of it afterwards. The work is in confirming that the quality holds on your code, and in making sure the model reaches your developers because you chose it.
What Microsoft Shipped in GitHub Copilot on November 17
Microsoft AI released two coding models at Ignite: MAI Code 2 for production work and MAI Code 2 Flash for high-volume, routine edits. Microsoft's launch post positions MAI Code 2 for deep debugging, multi-file implementation, repo-wide reasoning and UI work, and Flash for code explanations, targeted edits, routine fixes and fast iteration. The same post says neither model is "yet tuned" for harder problems such as very long-horizon agentic work, which matters if your main Copilot use is the cloud agent working unattended on a large ticket.
Both models went live in GitHub Copilot the same day. GitHub's changelog lists MAI Code 2 on Pro, Pro+, Max, Business and Enterprise, and Flash on every plan from Free upward, in the model picker in VS Code, Visual Studio, JetBrains, Xcode, Eclipse, Copilot CLI, the Copilot cloud agent, the Copilot app, github.com and GitHub Mobile. Both are billed under usage-based billing.
They were part of a wider release. Thurrott's Ignite roundup counts seven new MAI models, reports that all of them run in production on Microsoft's Maia 200 accelerator, and quotes Microsoft AI CEO Mustafa Suleyman: "We co-optimize quality and price together." Running on its own silicon is how Microsoft can price this low, and it is also why the price is Microsoft's to change. We covered the move of Copilot inference onto Maia in Microsoft Moves Foundry and GitHub Copilot to Maia 200 by Year-End.
What Does MAI Code 2 Cost Now and After May 17, 2027?
MAI Code 2 costs $0.50 per million input tokens and $2.50 per million output tokens until May 17, 2027, then $1.00 and $5.00. Those rates come from GitHub's models and pricing reference, checked on November 18, 2026, which lists the promotion and the post-promotion rate side by side.
| Model (GitHub Copilot, per 1M tokens) | Input | Cached input | Output |
|---|---|---|---|
| MAI Code 2, promo to May 17, 2027 | $0.50 | $0.025 | $2.50 |
| MAI Code 2, after the promo | $1.00 | $0.05 | $5.00 |
| MAI Code 2 Flash, promo to May 17, 2027 | $0.05 | $0.005 | $0.25 |
| MAI Code 2 Flash, after the promo | $0.10 | $0.01 | $0.50 |
| MAI-Code-1.1-Flash (retiring December 18) | $0.20 | $0.02 | $1.20 |
| GPT-6.1 Sol (up to 272K) | $2.00 | $0.10 | $10.00 |
| Claude Opus 5.5 | $4.00 | $0.20 | $20.00 |
| GPT-6 Luna (up to 272K) | $0.10 | $0.01 | $0.50 |
| Claude Haiku 5.5 (up to 100K) | $0.10 | $0.01 | $0.50 |
Prices from GitHub's models and pricing page, checked November 18, 2026.
Microsoft's cost ratios come from its benchmark runs (the post does not define the unit, and names a cost source only for FrontierCode), and its footnotes say MAI's cost is shown at the 50% promotional price. A run that uses the same tokens costs twice as much once the rate doubles. Apply that to Microsoft's own figures and the comparison moves:
- MAI Code 2 against GPT-6.1 Sol goes from about 40% of the cost to about 80%.
- MAI Code 2 against Claude Opus 5.5 goes from about 5% to about 10%.
- MAI Code 2 Flash against GPT-6 Luna goes from 65% of the cost to about 130%, which makes it more expensive per task than Luna by Microsoft's own numbers. Against Claude Haiku 5.5 it goes from 20% to about 40%.
That arithmetic assumes the model's token usage stays the same, and it uses Microsoft's ratios, which you have not yet reproduced. Even so, the Flash comparison should change how you plan. After May 17, Flash's per-token rate is identical to Luna's and Haiku 5.5's short-context rate, so "use Flash because it is cheapest" stops being automatic. MAI Code 2 still undercuts GPT-6.1 Sol and Opus 5.5 at list price.
What That Means for a Copilot Business Pool
A Copilot Business seat includes 1,900 AI credits a month (1 credit is $0.01) and Enterprise includes 3,900, pooled across the billing entity, per GitHub's usage-based billing docs. GitHub set those allowances at the $19 and $39 seat prices when it moved Copilot to usage-based billing on June 1, 2026.
Take an illustrative agent task that reads 1 million uncached input tokens and writes 100,000 output tokens. On list rates it costs $0.75 on MAI Code 2 during the promotion, $1.50 after it, $3.00 on GPT-6.1 Sol and $6.00 on Opus 5.5. One Business seat's $19 covers about 25 of those on promotional MAI Code 2, about 12 after May, about 6 on GPT-6.1 Sol and about 3 on Opus 5.5. Real tasks cache heavily and models use different token counts, so treat this as the shape of the curve. Your own billing export will give you the actual figure. For how per-developer Copilot bills build up under this model, see Your Copilot Bill Can Reach $137 a Month at 10 Daily Tasks.
Microsoft's 98% Claim Comes From Its Own Harness at Max Effort
The "98% of GPT-6.1 Sol" and "96% of Claude Opus 5.5" figures are an average of four benchmarks, three of which Microsoft ran itself. Microsoft's footnotes say DeepSWE 1.1, Terminal Bench 2.1 and SWE-Atlas were run on MSBench, Microsoft's internal evaluation platform, using the GitHub Copilot CLI without subagents, with every model at maximum reasoning effort. FrontierCode v1.1 scores and costs came from Artificial Analysis. The frontend result (reviewers preferred MAI Code 2's output over GPT-6 Sol 64% of the time) came from blind side-by-side reviews by 258 annotators.
Three things in that setup may not match how your teams work:
- Reasoning effort. Max effort is the most expensive setting and rarely what developers leave the picker on. Quality and cost both move with effort, and they may not move equally across models. We saw the same effect on another launch in Grok 4.7 Kept Its $2/$6 Price and Can Still Double Cost per Task.
- The harness. The numbers come from Copilot CLI with no subagents. A developer in VS Code agent mode, or the cloud agent splitting work, runs a different loop, and harness differences can reorder model rankings, as we covered in Your Bake-Off Ranked Harnesses, Not Models. Re-Run It.
- The task mix. Public benchmarks lean toward self-contained repos and Linux terminals. Microsoft's own caveat is that results "vary by workload, configuration, reasoning effort, and pricing assumptions."
The strongest case for the claim is that Microsoft ran its competitors through the same harness under the same settings, so the ratio is internally consistent. That is good evidence the models are close on those four benchmarks. How close they are on your monorepo, with your build system and your review standards, is something only your own run will show.
Is MAI Code 2 Already On for Your Developers?
Most likely, yes. GitHub's launch changelog says that under default model enablement "new models are enabled automatically unless an administrator has turned off the global default." GitHub turned that policy on for Business and Enterprise from August 26, per its July changelog. Any generally available model you never configured follows the policy, and with the policy enabled it is available to your users.
GitHub's default model availability docs name the control "Default availability for released models" and mark models that follow it as "Delegate to Default Policy." The same page lists what never switches on automatically: pre-GA models, open-weight models such as DeepSeek and Kimi K3, and models outside GitHub's data retention agreement, currently Claude Fable 5, 5.1 and 5.2. MAI Code 2 is in none of those groups.
The problem with a cheap model showing up in the picker is that it reached your developers without anyone deciding it should. On top of that, the "AI credits paid usage" policy is also enabled by default for Business and Enterprise, according to the billing docs. If a team standardises on MAI Code 2 Flash in December because it is the cheapest option, its per-task cost against Luna roughly doubles in May without anyone having made a new choice.
How to Run Your Own MAI Code 2 Evaluation
A coding model evaluation is a fixed set of your own past tasks, run through each candidate model in the harness your developers use, scored on whether the change would have merged and on what it cost. You can build a useful one in a week from work your teams have already closed.
- Pick 30 to 50 closed pull requests across your main languages and repos, weighted toward the task types you want MAI Code 2 for: multi-file changes, bug fixes and UI work. Leave out long unattended agent runs, since Microsoft says the model is not yet tuned for those.
- Run them in the surface your developers use, whether that is VS Code agent mode, Copilot CLI or the cloud agent, at the reasoning effort they leave the picker on. Run your current default model on the same tasks.
- Score merge-readiness as well as test results. Have a reviewer grade each diff as mergeable, mergeable after edits, or rejected, without knowing which model wrote it.
- Record cost per task from the billing export, then double the MAI figures to see the post-May cost. Compare cost per mergeable change, since a cheaper model that needs a second attempt can cost more.
- Repeat the run for Flash against GPT-6 Luna and Claude Haiku 5.5 on the small edits and explanations Flash is meant for. At list price all three charge the same per token, so the decision turns on quality and token count.
If MAI Code 2 holds up within a few points of your current default, the post-promo price still saves money against GPT-6.1 Sol and Opus 5.5. If it does not, the promotion is not a reason to move. For how per-phase routing changes the cost picture inside Copilot, see GitHub's Router Bills Every Leg. Find Your Break-Even.
MAI-Code-1.1-Flash Retires on December 18
GitHub is retiring MAI-Code-1.1-Flash across all Copilot experiences on December 18, 2026, per its deprecation notice, and points users to the latest MAI Code Flash model. Users and integration owners need to update workflows before that date, and Copilot Enterprise administrators need to enable the replacement through model policies in Copilot settings if it is not already enabled.
The replacement is cheaper. MAI-Code-1.1-Flash lists at $0.20 input and $1.20 output per million tokens, against $0.10 and $0.50 for MAI Code 2 Flash even after its promotion ends, per GitHub's pricing page. The risk is in anything that names the old model explicitly: a CLI script, an Actions workflow, a custom agent or an internal tool calling Copilot with a pinned model string. Those break on December 18.
What to Do This Week, This Month and Before May 17, 2027
This Week:
- Open the model policy in Copilot settings and check whether MAI Code 2 and MAI Code 2 Flash show "Delegate to Default Policy." Then decide, with your security lead, whether each one is enabled, disabled or limited to a pilot team. Write the decision down.
- Search your repos, Actions workflows and agent configs for the string "MAI-Code-1.1-Flash" and assign an owner to each hit, with December 18 as the deadline.
- Confirm whether the "AI credits paid usage" policy is on, and set user-level or cost-center budgets if it is.
This Month:
- Run the 30 to 50 task evaluation above against your current default, at your developers' real reasoning effort, and record cost per mergeable change at both the promotional and the list price.
- If MAI Code 2 holds up, make it the default for the task types where it scored well and keep your current model available for long agent runs.
- Post the decision in your engineering channel, with the May 17 date, so teams do not standardise on Flash because it is cheapest this quarter.
Before May 17, 2027:
- Re-price your Copilot forecast for the second half of your fiscal year at the post-promo rates: $1.00/$5.00 for MAI Code 2 and $0.10/$0.50 for Flash.
- Re-run the Flash evaluation against GPT-6 Luna and Claude Haiku 5.5, since all three cost the same per token at list.
- Ask your GitHub account team in writing whether the promotion will be extended, and whether a committed-spend agreement can lock in a rate.
The Bottom Line
Microsoft keeps shipping AI into tenants as on-by-default. It set Claude to all users in untouched UK Copilot tenants, GitHub folded Copilot into GitHub Enterprise, and MAI Code 2 arrived through default model enablement with a six-month discount attached. Introductory cloud pricing ran the same way: teams built on the launch rate and found out the real rate at renewal.
MAI Code 2 may well earn the default slot, and at list price it still costs half of GPT-6.1 Sol per token. Put your own evaluation and the May 17 date on paper before your developers pick it for you.
Continue Reading
- Your Copilot Bill Can Reach $137 a Month at 10 Daily Tasks
- GitHub Enterprise Adds Copilot, Bills From First Use
- Your Bake-Off Ranked Harnesses, Not Models. Re-Run It.
- GitHub's Router Bills Every Leg. Find Your Break-Even.
- GPT-6.2 Astra Costs 20% Less Than GPT-6 Astra on API Tiers
- Microsoft Adds Hard Spending Caps for Copilot Studio Agents
