MAI Code 2 Ships in GitHub Copilot: 98% of GPT-6.1 Sol at 40% of the Cost
Microsoft put its in-house MAI Code 2 coding models into GitHub Copilot and Code 2 Flash into Foundry and OpenRouter, claiming 98% of GPT-6.1 Sol's averaged benchmark performance at 40% of the cost, with a 50% promotional rate in the fine print.
Microsoft shipped MAI Code 2 and Code 2 Flash, its in-house coding models, into GitHub Copilot this week, with Flash also landing in limited preview in Microsoft Foundry and OpenRouter. The headline numbers are aggressive: averaged over four benchmarks, Microsoft claims 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 5% of the cost. Both cost figures carry a caveat that matters more than the percentages, because the 5% comparison is calculated at a 50% promotional rate.
The strategic read is simpler than the benchmark table. Microsoft spent years routing Copilot traffic to OpenAI models. With MAI Code 2, the company now owns a coding model it claims lands within 2% of GPT-6.1 Sol on averaged benchmarks while costing less than half as much to serve. For an enterprise that already pays for Copilot across thousands of seats, the model swap happens underneath them. The procurement decision was made years ago; the unit-economics decision is being made now, by Microsoft, on their behalf.
Key Takeaways: What Matters Most?
- Averaged over four benchmarks, Microsoft says MAI Code 2 delivers 98% of GPT-6.1 Sol's performance at 40% of the cost.
- Against Claude Opus 5.5, Microsoft claims 96% of the performance at 5% of the cost, at a 50% promotional rate.
- Code 2 Flash is in limited preview in Microsoft Foundry and OpenRouter, giving non-Copilot shops a way to test the model.
- Microsoft itself says MAI Code 2 is not yet tuned for very long-horizon agentic work, the exact pattern most enterprises are wiring up.
- Microsoft's MAI Work Flash already backs Microsoft 365 apps and FabricIQ, so this is one model family spreading across the stack.
Takeaways like these, every weekday morning. Get the briefing
The claim, and what it does not measure
Microsoft's numbers come from averaging four benchmarks, per the company's announcement. Averaging is a defensible way to summarize, but it hides the shape of the distribution. A model can sit at 99% on three benchmarks and 92% on the fourth and still average out to a flattering figure. What the 98% figure cannot tell you is whether that fourth benchmark is the one your codebase resembles. Vendor-averaged performance has never been a reliable predictor of per-repository performance, and nothing in this release changes that.
The 5%-of-Opus figure deserves extra scrutiny because it stacks two multipliers. The promotional rate is 50% off, and Claude Opus 5.5 is a frontier-class model priced accordingly. Put those together and you get a headline that is technically accurate and operationally misleading if you quote it in a budget deck without the asterisk. If your engineering org standardizes on this model and the promotion ends, your cost model changes the day Microsoft's pricing page does.
There is also an honest limitation in the announcement, and Microsoft deserves credit for stating it plainly: the company says MAI Code 2 is not yet tuned for very long-horizon agentic work. That is the exact workload category most enterprises are building toward, the multi-hour autonomous session that plans, edits, tests, and iterates across a repository. A model that is strong on single-shot and short-loop coding tasks but untested on long-horizon agency is not yet a drop-in replacement for the frontier models that are.
One model family, spreading across the stack
The release also clarifies Microsoft's architecture. MAI Work Flash already serves Microsoft 365 apps and FabricIQ, per the announcement. MAI Code 2 serves Copilot, Foundry, and OpenRouter. That is one in-house model family spreading across productivity, data analytics, and code, which changes the negotiation dynamics for every Microsoft customer. When the same model backs the office suite and the IDE, the question stops being "which coding model should we buy" and becomes "what is inside the E5 license we already pay for."
For OpenRouter users, the limited preview of Code 2 Flash is the most useful part of the news. It gives teams a way to run the model against their own eval harness without a Copilot commitment, side by side with GPT-6.1 Sol, Claude Opus 5.5, and Gemini 4 Neon. The week's other releases, including GPT-6.2 Astra's 20% price cut and Gemini 4 Neon at $3/$15 per million tokens, confirm that coding-model pricing is compressing across the board. MAI Code 2's pitch lands in a market where every vendor is already cutting prices, which is why the promotional-rate fine print matters.
The trade-off, stated plainly
The cost of adopting MAI Code 2 as a default is not measured in tokens. It is measured in dependence on a single vendor's roadmap. Microsoft is the only party that has run the four-benchmark average it published. If the model underperforms on your workload, you find out in your own review latency and your own regression counts, not in a report. And because the Copilot swap happens server-side, your team's experience can shift without any change you approved.
The second cost is benchmark myopia. Averaged benchmark performance at a discount is a procurement argument, not an engineering argument. The engineering argument needs per-task data: acceptance rate on your codebase, edit distance on your tests, behavior on migrations and multi-file refactors. The teams that will benefit from this release are the ones that run that eval before the promo window closes, not the ones that forward the benchmark chart to finance.
What To Do About It
- Run your own eval this week: Pipe 50 representative tasks from your backlog through Code 2 Flash on OpenRouter and compare acceptance rate against your incumbent model before quoting Microsoft's 98% figure internally.
- Price the steady state, not the promo: Rebuild your token-cost model at the non-promotional rate; if the 5%-of-Opus claim halves when the 50% promo ends, your savings case changes materially.
- Keep agents where they work: Leave long-horizon autonomous coding sessions on your proven frontier model until Microsoft ships tuning data for agentic workloads, per its own caveat.
- Check your Copilot seat math: If MAI Code 2 is now serving your Copilot traffic, measure review-time deltas this sprint so you notice quality shifts the day they happen, not a quarter later.
The Bottom Line
MAI Code 2's numbers make Copilot the cheapest credible default in enterprise code, if the 98%-of-Sol average survives contact with your repository. The promotional rate and the long-horizon caveat are not footnotes; they are the two facts that decide whether this saves you 60% or 5% a year. Test the model on your own tasks now, while the preview is open and the promo rate holds, because the pricing leverage you have this month is the pricing leverage you will not have once the standard rate applies.
Sources
Frequently Asked Questions
Share this article