Brief IA
›

Microsoft Unveils Two Coding Models at Ignite 2026

Code & Dev·Tom Levy·

Microsoft Unveils Two Coding Models at Ignite 2026

Microsoft Unveils Two Coding Models at Ignite 2026
Key Takeaways
1Microsoft launched MAI Code 2 and MAI Code 2 Flash at Ignite 2026.
2The comparative costs of its models include a 50% promotion.
3Microsoft acknowledges that these models are not yet optimized for long-term agentic work.
Why it matters : Microsoft is developing its own AI models as part of its efforts to reduce reliance on external vendors.

Microsoft has launched MAI Code 2 and MAI Code 2 Flash, two coding models presented at Ignite 2026. The company claims they can compete with offerings from OpenAI and Anthropic. However, its cost comparisons are based on promotional rates.

Other Costs of Software Development

Integration, testing, security checks, human verification, and fixing errors produced by AI are also part of the expenses, Jain reminds us. His example is conditional: if the models account for 10% of the development budget, reducing their cost by 60% would only yield a 6% savings overall. He therefore encourages considering the cost of delivering functional software rather than just the price of generating code.

Sanchit Vir Gogia, chief analyst at Greyhound Research, advocates for testing with the company's tools and code on production work. He recommends comparing the cost of each merged change, the failures of changes, and security outcomes to those of teams without AI coding models. He also clarifies that the amounts compared by Microsoft correspond to running an assessment, not a token pricing grid, and believes that the pricing advantage diminishes significantly after the promotion.

Gogia reminds us that Google's DORA research has linked increased AI usage to lower delivery stability, with potential hidden costs. Additionally, Microsoft assesses the offensive cybersecurity capabilities of its two models at a high level. For Gogia, CIOs must therefore measure not only code quality but also its security and reliability after production integration.

GitHub Displays Temporarily Reduced Prices

Microsoft estimates the average task cost at $0.309 with MAI Code 2, compared to $0.761 with GPT-6.1 Sol and $6.306 with Claude Opus 5.5. For the Flash variant, the company announces $0.052, against $0.080 for GPT-6 Luna and $0.319 for Claude Haiku 5.5. The amounts retained for its own models include a 50% promotional discount.

La lettre du soirGratuit · 19 h

Le brief IA que les pros lisent chaque soir

Les 7 actus IA du jour, décryptées en 5 min. Gratuit.

Inclus dès l'inscription : notre sélection des meilleurs guides & comparatifs IA.

Choisissez votre rythme

Gratuit · Pas de spam · Désabonnement en 1 clic

GitHub displays, until May 17, 2027, the following rates for MAI Code 2: $0.50 per million tokens received as input, $0.025 for cached inputs, and $2.50 for output tokens. These prices will then rise to $1.00, $0.05, and $5.00, respectively.

For MAI Code 2 Flash, the promotional rates per million tokens are $0.05 for input, $0.005 for cached inputs, and $0.25 for output. The amounts announced after the promotion are $0.10, $0.01, and $0.50, respectively.

Two Models for Common Coding Tasks

According to Microsoft, the MAI Code 2 version is intended for advanced debugging, changes spread across multiple files, reasoning over an entire repository, and creating user interfaces. The Flash variant is aimed more at high volumes: code explanation, targeted tweaks, routine fixes, and rapid iteration cycles.

The company claims to have prioritized tasks that its clients perform every day. It acknowledges that these models are not yet optimized for fundamental scientific research or for long-duration agentic activities. Their development is part of Microsoft's efforts to have proprietary capabilities and to rely less on external suppliers.

Variable Results Based on Coding Tests

For MAI Code 2, Microsoft uses DeepSWE 1.1, Terminal-Bench 2.1, SWE-Atlas, and FrontierCode v1.1, test benches employed in GitHub Copilot. The company reports an average success rate of 63.2%, behind GPT-6.1 Sol at 64.1% and Claude Opus 5.5 at 65.3%.

On DeepSWE 1.1, the scores reported by Microsoft are 73.2% for its model, 74.6% for OpenAI's, and 69% for Anthropic's. On Terminal-Bench 2.1, they reach 84.6%, 85%, and 78.7%, respectively. SWE-Atlas gives, in this same order, 43.1%, 47.4%, and 57.5%. On FrontierCode v1.1, Microsoft indicates 52%, 49.4%, and 56%.

The comparison of Flash with GPT-6 Luna and Claude Haiku 5.5 is based on DeepSWE 1.1, Terminal-Bench 2.1, and SWE-Atlas, at the maximum reasoning level. This time, Microsoft announces an average of 61% for Flash, ahead of 55.1% for GPT-6 Luna and 53.9% for Claude Haiku 5.5.

In the detailed results from Microsoft, DeepSWE 1.1 gives 59.9% to Flash, 59.6% to GPT-6 Luna, and 46% to Claude Haiku 5.5. Terminal-Bench 2.1 assigns 78.3%, 69.3%, and 73.4% to these models, respectively. For SWE-Atlas, the reported scores are, in the same order, 44.8%, 36.4%, and 42.1%.

Brief IA — L'actualité IA en français

L'essentiel de l'actualité de l'intelligence artificielle, décrypté et expliqué chaque jour.