Copilot for three years has been a Microsoft product running on OpenAI's models. Build 2026, which opened this morning in San Francisco, is the announcement that puts a clock on that sentence.

Project Polaris is Microsoft's answer. In-house mixture-of-experts coding model. Runs on custom Maia AI accelerators inside Azure. Targeting GitHub Copilot GA by August 2026. Per Microsoft's internal benchmarks, it beats GPT-4 Turbo on HumanEval and MBPP — the field-standard code generation evaluations — with the largest gains in low-resource languages like Rust and Haskell.

Those benchmarks are Microsoft's. No independent reproduction exists at the time I'm writing this. First-party numbers on an unshipped model should be treated as marketing until someone outside the company confirms them. I'll keep saying that.

What I don't need third-party confirmation for: Polaris runs on Maia accelerators, inside Azure, on infrastructure Microsoft controls. That's an inference stack they own end-to-end. Real, regardless of whether the HumanEval number holds.

Polaris isn't the only announcement. Microsoft shipped MAI-Image 2.5, MAI-Voice 2, MAI-Transcribe 1.5, and MAI-Thinking-1 alongside it. That last one is the entry worth sitting with. MAI-Thinking-1 is Microsoft's first dedicated reasoning model. You don't build a reasoning model if you plan to keep buying reasoning from somewhere else.

Copilot was rearchitected as a multi-model platform. It now routes tasks across OpenAI, Anthropic, and open-source models depending on the workload. The framing is careful — Microsoft didn't announce "Copilot now runs on Polaris." They announced "Copilot routes to the best model for the job." Polaris is one option. OpenAI is still in there. So is Anthropic.

And Windows Agent Framework v1.0 shipped as a production MIT-licensed release — the formal convergence of AutoGen and Semantic Kernel into one SDK. Agents defined in YAML. Portable across local Windows, Windows 365, and Azure without changing the manifest. Cross-agent communication over gRPC. Memory service for persistent caching. The kind of infrastructure layer that takes eighteen months to appreciate.

Source spread

• AI Weekly — Microsoft targets Claude Code with Project Polaris — [builder]. Best on the August GA date, HumanEval/MBPP claims, and competitive framing against Claude Code.
• Windows News — Build 2026: homegrown models — [builder]. Technical coverage of Polaris architecture; good on the full MAI family.
• TechCrunch — 20M+ paid Copilot users (April 2026) — [builder]. The user count that makes the volume of this inference shift legible.
• Om Malik — What Microsoft's 10-Q says about OpenAI — [skeptic]. The exclusivity clause removal buried in a quarterly filing; essential context.
• Microsoft Agent Framework blog — v1.0 — [builder]. Primary source on WAF production release and MIT licensing.

Pros & cons

What's real:

• Polaris on Maia accelerators inside Azure means Microsoft controls latency, pricing, and data residency for its most-used developer product. These operational advantages compound over time and don't show up in a benchmark.
• Multi-model Copilot routing is architecturally correct. Different workloads have different cost and quality tradeoffs. Builders get routing flexibility they didn't have before.
• MAI-Thinking-1 existing at all is the signal. You build a dedicated reasoning model when you're serious about the full stack — not just the code-completion tier.
• WAF v1.0 being MIT-licensed means commercial production use is unambiguous. The YAML-based agent portability pattern is worth learning before the GA ecosystem hardens around it.

What deserves a side-eye:

• The benchmark claim — "beats GPT-4 Turbo on HumanEval and MBPP" — is Microsoft's internal assessment on a model not yet in GA. No SWE-bench Verified. No TAU-bench. No third-party reproduction. Don't make production architecture decisions based on these numbers yet.
• Polaris doesn't reach GA until August. A Build announcement and a shipping model are two different things, and the gap between them has historically included surprises.
• Multi-model Copilot routing adds debugging surface area. If the routing logic changes which model handles your task, answers change in ways you may not have opted into.

Further reading

• Microsoft Agent Framework v1.0 — devblogs.microsoft.com — primary source on architecture and MIT licensing
• AI Weekly — Microsoft targets Claude Code with Project Polaris
• Windows News — Build 2026: homegrown models for Copilot
• Om Malik — What Microsoft's 10-Q says about OpenAI
• TechCrunch — Microsoft says it has over 20M paid Copilot users