98% of the Performance at 40% of the Price - But Three of Four Benchmarks Ran In-House: Microsoft Puts MAI Code 2 Inside Copilot
98% of the Performance at 40% of the Price — But Three of Four Benchmarks Ran In-House: Microsoft Puts MAI Code 2 Inside Copilot
TL;DR: What actually shipped
The short version: Microsoft has put two of its own coding models directly into Copilot's model picker and is using a very low sticker price to win the default slot.
On November 17, 2026, at its Ignite conference, Microsoft added two coding models to GitHub Copilot: MAI Code 2 and MAI Code 2 Flash. They are the third generation of the in-house MAI-Code family, and the first ones positioned to move beyond lightweight completion into full agentic software engineering.
The division of labor is simple:
- MAI Code 2 handles the heavy work: multi-file changes, deep debugging, repository-wide reasoning, and interface building.
- MAI Code 2 Flash handles high-volume light work: explaining code, targeted edits, routine fixes, and fast iteration.
What matters here is not "another model appeared." It is that Microsoft is now selling its own model next to OpenAI's and Anthropic's inside its own distribution channel. Price is the weapon, and price is also the biggest point of dispute.
This article is compiled from official changelogs, official pricing documentation, and multiple media reports.
Pricing: startlingly cheap, with two conditions
GitHub's pricing documentation is blunt about it. Both models are usage-based, and both are currently on promotional pricing:
- MAI Code 2: $0.50 per million input tokens, $0.025 cached input, $2.50 output.
- MAI Code 2 Flash: $0.05 per million input tokens, $0.005 cached input, $0.25 output.
The gap is easy to feel: Flash's input rate is one tenth of Code 2's, and its output rate is also one tenth.
But keep two things in mind, because this is where teams get burned:
- This is promotional pricing, and every rate doubles on May 17, 2027. Code 2 becomes $1.00 / $0.05 / $5.00; Flash becomes $0.10 / $0.01 / $0.50. If you build an annual budget on today's numbers, that budget doubles next May.
- You pay per token, not per request. Let a task loop a few extra turns or read a few extra files, and the bill follows. What is cheap is the unit rate, not your monthly total.
The official scorecard: where 98% and 96% come from
Microsoft's headline numbers are eye-catching. Averaged across four benchmarks, MAI Code 2 delivers 98% of GPT-6.1 Sol's performance at 40% of the cost, and 96% of Claude Opus 5.5's performance at just 5% of the cost.
The four benchmarks are DeepSWE 1.1, Terminal Bench 2.1, SWE-Atlas, and FrontierCode v1.1. All of them target the kind of coding work enterprises actually see day to day, not contest problems.
Flash gets a similar pitch: across three benchmarks measuring high-volume routine coding, it performs comparably to GPT-6 Luna and Claude Haiku 5.5 at 65% and 20% of their cost respectively.
One more number deserves its own line. In blind frontend evaluations covering interfaces, small games, data visualizations, 3D scenes, and SVG graphics, developers preferred MAI Code 2's output over GPT-6 Sol 64% of the time. Aesthetic and functional quality were scored together, by 258 annotators in total.
But those numbers come with caveats, and Microsoft wrote them in the footnotes
This is the part worth slowing down for. Microsoft's own footnotes are candid:
- Three of the four benchmarks (DeepSWE 1.1, Terminal Bench 2.1, SWE-Atlas) ran on Microsoft's internal MSBench evaluation platform, using the GitHub Copilot CLI, without subagents.
- FrontierCode 1.1 scores and costs came from the third-party Artificial Analysis.
- Every model ran at maximum reasoning effort.
- Cost comparisons used the 50% promotional price.
Put those four together and the picture sharpens. "40% of the cost" compares a promotional rate against competitors' list prices. "98% of the performance" comes mostly from Microsoft's own harness, single-process, without subagents. That is not deception, but it is also not the same thing as what you will see in your own repository.
The press read it the same way. InfoWorld framed the launch on November 18 as a step toward reducing Microsoft's dependence on OpenAI and Anthropic. Crypto Briefing noted the same day that independent benchmarks have not verified the performance claims, and pointed out that one of the standout data points came from an internal AT&T evaluation against certain GPT-6.1 variants rather than the whole family. Techzine's headline was simply that the model generates code at a fraction of the price.
Who can use it, and where
The rollout is broader than you might expect:
- MAI Code 2: Copilot Pro, Pro+, Max, Business, and Enterprise.
- MAI Code 2 Flash: all of the above, plus Copilot Free and Student.
That means free users get Flash too. This matters more in practice than any benchmark score, because it makes "try Microsoft's own model at zero cost" a real option.
The entry points are extensive: VS Code, Visual Studio, Copilot CLI, the Copilot cloud agent, the GitHub Copilot app, github.com, GitHub Mobile on iOS and Android, JetBrains, Xcode, and Eclipse. Enterprise administrators can toggle both models through model policies in Copilot settings.
One deadline to note: MAI-Code-1.1-Flash is retired across all Copilot experiences on December 18, 2026, and the suggested replacement is the new Flash. If your team has scripts or CI configuration with the old model name hard-coded, that work should start now.
To be clear about what kind of product this is: Copilot is a closed-source commercial subscription, and neither new model is released as open weights. You can only call them through Copilot or Microsoft's own channels under usage-based billing. This launch is about the model lineup inside a subscription, not about open-source tooling.
How to use it: three steps, and don't switch everything at once
- Change the model first, not the workflow. Pick MAI Code 2 or Flash in Copilot's model picker and run the one thing you do most often — explaining unfamiliar code, or fixing a lint error. This step only answers "is it good enough," not "does it save money."
- Then read the billing units. Open the Copilot usage page and check whether you are spending input or output tokens. Output costs five times input on Code 2 ($2.50 versus $0.50), so "let the model write more" is far more expensive than "let the model read more."
- Set policy last. In enterprise settings, admins should confirm in model policies whether both models are enabled for the relevant teams. Individual users should trial within their Pro or Pro+ allowance in small batches rather than changing the default model outright.
Good fit, and not a good fit
Good fit:
- Mid-to-large projects that involve constant edits across many files, where Code 2's repository-wide reasoning is the main draw.
- High-volume, repetitive, low-risk changes such as adding tests, fixing types, or clearing lint warnings, where Flash's unit price advantage is clearest.
- Budget-sensitive teams that do not want to abandon the Copilot workflow, especially organizations paying per seat with heavy code volume.
- Anyone who wants a free way to try Microsoft's in-house model, since Flash is available on Copilot Free.
Not a good fit:
- Very long-horizon agentic work. Microsoft states in its own announcement that these models are "not yet tuned" for fundamental science research or very long-horizon agentic work. Do not expect them to hold up over multi-hour autonomous runs.
- Teams that need independent third-party benchmark backing before procurement. Current results are largely vendor-run.
- Workflows deeply committed to Claude Opus 5.5 or GPT-6.1 Sol with zero tolerance for quality drift. The hidden cost of switching — re-tuning prompts, rebuilding trust — often exceeds the savings.
Pitfalls to watch
- Do not treat the promotional price as the long-term price. Both tiers double on May 17, 2027. Run your budget at the doubled rate as well.
- Do not read "98%" as "98% in your scenario." Three benchmarks ran on Microsoft's own platform, single-process, without subagents. That is not the same as your IDE with a pile of tool calls attached.
- Do not miss December 18. MAI-Code-1.1-Flash goes away, and old configurations will fail silently.
- Flash in Microsoft Foundry is a limited-access preview, not general availability. Confirm access before planning around it.
- The "coming soon" listings for OpenRouter, Vercel, Hermes, OpenCode, and Kilo have no date. Microsoft says only "soon." Do not put them on a project schedule.
- There was an availability incident on November 18 at 09:34 UTC. GitHub Status reported an upstream model provider issue that was later resolved. Early in a launch, expect some of this.
Cost and time estimates
Rough math at promotional rates: a typical medium-sized change, assuming 200,000 input tokens (reading files and context) and 20,000 output tokens.
- MAI Code 2: 200,000 × $0.50/M = $0.10; 20,000 × $2.50/M = $0.05; about $0.15 total.
- MAI Code 2 Flash: 200,000 × $0.05/M = $0.01; 20,000 × $0.25/M = $0.005; about $0.015 total.
That is a tenfold gap. But if the task retries five times, Code 2 costs $0.75 and Flash costs $0.075. What determines the bill is retry count, not unit price.
On latency, Microsoft published no figures. Real-world speed depends on your reasoning-effort setting and task complexity, and the only way to know is to try it.
What to do next
- Today: switch to MAI Code 2 Flash in the Copilot model picker and run three familiar daily tasks on a Free or Pro account. Record how often the output was usable as-is.
- This week: take your most expensive workflow — usually cross-file refactoring — and run it once with Code 2. Compare actual time and rework rate against your current model.
- Before month end: if you decide to adopt, put the May 17, 2027 price increase into the budget, and file the December 18 retirement as a CI change ticket.
The real move here is not "Microsoft shipped a cheap model." It is that Microsoft put its own model on its own distribution channel and is using price to contest Copilot's default choice. For individual developers, that means one more option that is free to try and extremely cheap per token. For buyers and team leads, it means one more calculation to run — the one after the promotional pricing expires.
Sources
- https://github.blog/changelog/2026-11-17-mai-code-2-and-mai-code-2-flash-in-github-copilot/
- https://microsoft.ai/news/introducing-mai-code-2-and-mai-code-2-flash/
- https://docs.github.com/en/copilot/reference/copilot-billing/models-and-pricing
- https://github.blog/changelog/2026-11-17-upcoming-deprecation-of-mai-code-1-1-flash-in-december/
- https://www.infoworld.com/article/4250152/microsoft-launches-mai-code-2-models-to-challenge-openai-and-anthropic-in-ai-coding.html
- https://cryptobriefing.com/microsoft-mai-code-2-flash-github-copilot/