AI Momentum 107.6 +2.7%GitHub Pulse 5172 -5.9%Capital Flow $204.6B +5804.2%Hiring Heatmap 0.0 -100.0%
Advertisement
Models & Research

Microsoft's new coding AI looks cheap mainly because of a half-price offer. The price doubles in May

Microsoft wants more of the coding work done in its products to run on its own AI models instead of OpenAI's or Anthropic's. The pitch for its newest coding model, MAI Code 2, rests on price. Much of that price advantage is temporary.

Launched on Tuesday at Microsoft's Ignite conference, MAI Code 2 and a smaller sibling, MAI Code 2 Flash, are now inside GitHub Copilot, Microsoft's paid coding assistant. Microsoft says MAI Code 2 delivers 96% of the performance of Anthropic's Claude Opus 5.5 at 5% of the cost. It also claims 98% of OpenAI's GPT-6.1 Sol at about 40% of the cost.

A footnote on Microsoft's own chart says its model's cost is shown at a 50% promotional price. Copilot's rate card, as reported by several outlets, lists MAI Code 2 at $0.50 per million input tokens and $2.50 per million output tokens until May 17, 2027. After that, both rates double, to $1 and $5. Tokens are the small chunks of text that AI services bill by.

Two days after launch, no outside tester has published its own score. Artificial Analysis, a widely followed independent testing firm, had no listing for MAI Code 2 as of Thursday. Every quality and cost figure so far comes from Microsoft.

What the savings look like at full price

Microsoft published an average cost per task across four coding tests. MAI Code 2 cost about 31 cents per task and passed 63.2% of tasks. GPT-6.1 Sol cost 76 cents and passed 64.1%. Opus 5.5 cost $6.31 and passed 65.3%. Every model was run at its highest reasoning setting, meaning it could work through each problem as long as it needed.

By our arithmetic, doubling Microsoft's figure to reflect list price puts MAI Code 2 at about 62 cents a task. That is roughly 81% of GPT-6.1 Sol's cost, not 40%. Against Opus 5.5 it rises to about 10%, not 5%. The model stays far cheaper than Anthropic's. Against OpenAI's, most of the saving disappears.

The reason is volume. Per token, MAI Code 2 is priced at a quarter of GPT-6.1 Sol during the promotion. But dividing Microsoft's cost-per-task figures by those prices implies it uses about 1.6 times as many tokens to finish a job. That is our estimate, not a Microsoft figure; several vendor blogs reached the same number.

The smaller Flash model has the same issue. Microsoft says it beats OpenAI's GPT-6 Luna on its tests at 65% of Luna's cost per task. Flash's list price also doubles in May. Microsoft's cost figures for its own models use the promotional price, so by our arithmetic Flash would cost roughly 30% more than Luna per task at full price.

Measured against rivals that have already moved on

Microsoft ran three of its four tests on its own internal testing system. Only one, FrontierCode, comes from an outside lab.

The rivals in its charts are also aging. OpenAI released GPT-6.2 Sol, the successor to GPT-6.1 Sol, the same day Microsoft announced MAI Code 2.

Microsoft itself flags limits. It says the models are not yet tuned for hard science problems or for very long jobs that run unattended. On SWE Atlas, one test of complex software work, MAI Code 2 trails Opus 5.5 by 14.4 points, according to an analysis of Microsoft's published scores.

Access is narrow, too. MAI Code 2 is available only inside Copilot's paid plans, with no direct way for other software to call it. Flash is also in Copilot's free and student tiers and in a limited preview on Microsoft's Foundry cloud service. Support in outside coding tools such as OpenRouter and Kilo is promised "soon," with no dates.

Why Microsoft is pushing its own models

The strategic goal is control over what Copilot costs Microsoft to run. A request answered by an in-house model is one Microsoft does not have to buy from a rival lab. Microsoft ended its exclusive arrangement with OpenAI earlier this year. It has since released its own models for coding, images, speech and security.

For corporate buyers there is a practical catch. GitHub switches new models on by default for business and enterprise accounts unless an administrator blocks them. Teams could start relying on MAI Code 2 without anyone deciding to, at a price set to double in six months.

Investor newsletter Finimize argued that cheaper models bundled into everyday tools put pricing pressure on standalone model makers such as OpenAI and Anthropic. That pressure holds only if the savings last. For now, buyers have Microsoft's word on quality and a discount with an end date. Buyers who try it on routine work during the promotion should re-check the numbers before May.