GitHub Copilot’s default model is changing this month, and it’s not a newer OpenAI model doing the replacing. Microsoft’s in-house coding model — widely reported under the codename “Project Polaris” since its unveiling at Build 2026 on June 2 — becomes the default reasoning engine for every Copilot subscriber starting in August 2026, ending the product’s reliance on GPT-4 Turbo.
Quick facts
- Microsoft’s in-house coding model replaces GPT-4 Turbo as GitHub Copilot’s default starting August 2026, migrating automatically for all subscribers.
- It’s a Mixture-of-Experts model with sub-modules specialized by programming language, using chain-of-thought and tree-of-thought reasoning for multi-file edits.
- Teams that want to stay on GPT-4 Turbo get an optional fallback window through November 2026 before automatic migration becomes permanent.
- Other models, including Claude, Gemini, and Grok, remain selectable inside Copilot; this changes the default, not the only option.
- Microsoft’s own benchmark claims, outperforming GPT-4 Turbo on HumanEval and MBPP, haven’t been independently verified as of this writing.
Why Microsoft built its own coding model
GitHub Copilot has run on OpenAI models since it launched, and that dependency has shaped the product’s economics from day one. The timing here isn’t incidental: Microsoft and OpenAI restructured their partnership on April 27, 2026, ending Azure’s exclusivity over OpenAI model distribution and letting OpenAI sell through AWS and Google Cloud as well. Building a proprietary coding model gives Microsoft control over its own roadmap for its highest-volume developer product, reduces per-token costs, and stops it from being undercut on pricing inside its own platform, while Microsoft retains an IP license to OpenAI’s models through 2032, so this is a rebalancing, not a breakup.
What the model actually does differently
The architecture is a Mixture-of-Experts design with sub-modules tuned to specific programming languages and frameworks, running on Microsoft’s own Maia AI accelerators inside Azure rather than third-party infrastructure. Microsoft says it’s particularly strong on low-resource languages like Rust and Haskell, where general-purpose models more often hallucinate APIs that don’t actually exist, a real, specific weak point in current-generation coding assistants. At inference time, the model uses chain-of-thought and tree-of-thought reasoning aimed at multi-file refactors, the category of task where simpler single-file suggestion models tend to fall apart.
It shipped alongside a second, arguably more consequential change: multi-agent mode in VS Code, now in public preview, which lets an orchestrator agent spawn parallel subagents that handle linting, testing, documentation, and security review simultaneously instead of one after another on the same session. That pairs directly with the model swap, since a faster, cheaper in-house model matters more once Copilot is running several agent sessions in parallel rather than one at a time.
What to actually do before the migration
The migration is automatic and requires no enrollment, and Microsoft isn’t changing plan pricing alongside it. If your team has workflows tuned specifically around GPT-4 Turbo’s behavior, particular prompt patterns, expected output formatting, known failure modes you’ve already built tooling around, the practical move is to test the new model against your actual codebase during the fallback window rather than waiting for the hard cutover. The fallback to GPT-4 Turbo runs through November 2026; after that, reverting requires switching to GPT-4 Turbo manually as a selectable model rather than getting it by default.
How this fits the broader coding-agent landscape
Copilot isn’t competing in a vacuum. On Terminal-Bench 2.1, independent tracking currently shows GPT-5.6 Sol at 89.5% and Claude Opus 5 close behind at 89.1%, with the two most widely used coding agents essentially tied on their default configurations. Against that backdrop, Microsoft controlling its own model, rather than depending on whichever frontier model OpenAI ships next, is as much a strategic hedge as a capability play. With roughly 4.7 million paid Copilot subscribers and reported adoption across the large majority of Fortune 500 companies, even a modest quality change at the model layer affects an enormous amount of code being written daily.
Common questions
Do I need to do anything right now? No. Migration is automatic. If you want to test before the switch fully lands, check your Copilot model settings for a preview toggle; otherwise it happens on its own.
Will this cost more? Microsoft says plan pricing isn’t changing alongside the model swap. Copilot’s broader move to usage-based billing via GitHub AI Credits happened separately, on June 1, 2026.
Can I keep using Claude or Gemini inside Copilot instead? Yes. This change is only about which model is selected by default; Copilot’s multi-model picker, including Claude, Gemini, and Grok, isn’t going away.
What happens after the November fallback window closes? Reverting to GPT-4 Turbo will require manually selecting it from the model picker each time, rather than getting it as your default.
Key takeaway
If you use GitHub Copilot and haven’t opted into anything, you’re getting a new default model this month whether you asked for it or not. The fallback to GPT-4 Turbo exists specifically so you don’t have to find out the hard way whether the new model handles your codebase as well as advertised, so use the window through November to actually check, rather than assuming Microsoft’s internal benchmarks translate to your specific stack.
Leave a Reply