Microsoft pushes cheaper coding models, with promotional prices doing part of the work
Microsoft’s new coding models divide complex engineering from routine fixes. Their pricing pitch comes with a discount deadline, while Copilot users also face an older model’s retirement. The meaningful comparison is total cost per completed task, including retries and human rework.
Microsoft is moving the coding-model contest toward the bill. On November 17 it said MAI Code 2 averaged 98% of GPT-6.1 Sol’s performance at 40% of the cost across four selected coding benchmarks. That comparison uses a 50% promotional price, not a promise of equivalent savings on every project. The models are not yet tuned for fundamental science or very long autonomous tasks.
GitHub describes a division of labor: Code 2 handles deep debugging and changes spanning multiple files, while Flash targets explanations, small fixes and frequent iteration. Paying the same premium for both kinds of work makes it harder to see where the money goes.
Both arrived in Copilot on November 17. Code 2 covers paid plans starting with Pro; Flash also reaches Free and Student users. Business administrators control access through model policies. Microsoft’s separate Foundry availability is narrower: only Flash is in limited-access preview.
The pricing table separates material sent in, reused cached context and generated output. Per million tokens, Code 2 costs $0.50 for input and $2.50 for output; Flash costs $0.05 and $0.25 respectively. These are usage rates, not subscription fees or a fixed charge for finishing a project.
The footnotes also give the discount an endpoint: the listed rates double from May 17, 2027. A useful comparison should therefore run the same jobs through competing systems and count completion, total consumption and human rework. A cheap unit can still produce an expensive unsuccessful attempt.
There is a near-term migration deadline too. GitHub says MAI-Code-1.1-Flash will retire on December 18 and recommends moving workflows to the latest Flash model. Teams using the old model need to check configuration and results before that switch, not merely add another name to their shortlist.
iHolyshit review What interests me is cheaper coding inside tools developers already use, rather than another leaderboard claim. Lower prices expand experimentation; dependable task completion earns a place in production. After the promotion, fewer retries and fewer escalations to expensive models will be the more revealing measures of progress.
Share this article
What do you think of this story?
Anonymous feedback, up to 5 reactions per article.
Reader comments
Comment without signing in, or sign in to display your member name.
Discuss the technology and the story. Hate, politics and war, threats, abuse, personal accusations, sexual content, ads, links and personal information are not allowed.
No comments yet. Be the first to share.