Project Polaris — Microsoft Just Cut the OpenAI Cord for Copilot

On June 1, Microsoft flipped every paid GitHub Copilot subscriber to usage-based AI Credits billing. Twenty-four hours later, at Build 2026 in San Francisco, they announced Project Polaris — an in-house coding model that replaces GPT-4 Turbo as the default for all Copilot tiers starting August 2026. Two moves in 48 hours: first break the flat-rate pricing that made Copilot feel cheap, then reroute every metered token through a model Microsoft owns outright instead of one it pays OpenAI for. That is not a product upgrade. That is vertical integration executed in sequence.

TL;DR

• What: Microsoft announced Project Polaris at Build 2026 — an in-house model replacing GPT-4 Turbo as Copilot’s default starting August 2026
• Compounding move: Usage-based AI Credits billing went live June 1, one day before the Polaris reveal — every token now costs money, and soon every token runs through Microsoft’s own model
• Risk: Copilot SDK teams have until August to validate prompt behavior against Polaris before automatic migration removes the choice
• Action: If you are on Copilot Business or Enterprise, start testing Polaris on critical workflows now — the three-month GPT-4 fallback is a grace period, not an exit

Project Polaris — What Happened

Microsoft revealed Project Polaris during the Build 2026 keynote on June 2. It is a mixture-of-experts coding model built in-house, trained to run on Microsoft’s custom Maia AI accelerators inside Azure. Starting August 2026, Polaris becomes the default model for every Copilot Pro, Pro+, Business, and Enterprise subscriber through automatic migration. Teams that want to keep GPT-4 Turbo can enable a three-month fallback window, but it requires explicit opt-in — it does not persist on its own.

Microsoft claims Polaris outperforms GPT-4 Turbo on HumanEval and MBPP, with particular gains on languages like Rust and Haskell. No independent SWE-bench Verified or Terminal-Bench scores were published at Build. The benchmarks are self-reported with no external validation. Take them as directional, not definitive.

The strategic context is hard to miss. GitHub Copilot has relied on OpenAI’s Codex and GPT-4 family since 2021. Every completion, every code review suggestion, every agentic workflow has been running through a model Microsoft pays OpenAI per-token to use. Polaris running on Maia accelerators eliminates that licensing cost entirely. This is the same playbook Google ran with its Gemini-native tooling and Anthropic ran by deprecating third-party agent subscriptions — own the model, own the margin.

The three-month GPT-4 Turbo fallback must be explicitly enabled by an admin. If your organization does nothing, you are on Polaris in August with no rollback path. Mark your calendar.

Why This Matters

The billing change compounds the model change in ways that are easy to underestimate. GitHub AI Credits are priced at 1 credit = $0.01 USD, metered per token at published model rates. On June 1, model multipliers shifted dramatically: Claude Opus 4.7 jumped from 7.5× to 27×, and GPT-5.3 Codex jumped from 1× to 6×. Annual plan subscribers who budgeted based on the old multipliers are already being squeezed before Polaris even lands.

Now layer the model switch on top. If you are running code review automation through Copilot, every review consumes both AI Credits and GitHub Actions minutes at standard per-minute rates. When Polaris becomes the default in August, the token pricing for that model will determine your actual cost — and Microsoft controls both the price and the model generating the tokens. There is no market pressure on Polaris pricing because there is no alternative provider for the same model. You pay what Microsoft charges or you leave the platform.

I have been tracking this trajectory since Anthropic’s SDK billing split and OpenAI’s Codex moves — every major vendor is trying to own the full stack from model to meter. Microsoft just did both in one week. The pattern repeats: train users to run expensive agentic workflows on flat-rate plans, then flip the pricing model before they have had time to build alternatives. What makes Polaris different from a normal model refresh is the lock-in geometry. If you are building internal tooling on the Copilot SDK, or running CI pipelines through Copilot Actions, you now have a vendor-controlled model baked into every layer of that stack. Your prompts were tuned against GPT-4 Turbo behavior. Your output parsers expect GPT-4 Turbo formatting. Your edge-case handling for less common languages was validated against GPT-4 Turbo responses. Polaris is a mixture-of-experts architecture — meaning outputs for edge-case languages may differ meaningfully from what your prompts were optimized for.

If your team uses Copilot SDK for internal tooling or CI automation, start a parallel validation branch against Polaris now. Focus on: prompt behavior consistency, output format stability, and refactoring accuracy for your specific language mix. Do not wait for August to discover regressions.

The competitive framing matters too. Claude Code overtook Copilot in developer adoption, and Microsoft is explicitly positioning Polaris as the response — competing on model quality and billing infrastructure simultaneously. But there is a pressure release valve developers are already finding: model-agnostic editors like Cursor that let you bring your own API key and pick your model per task. When a vendor locks you into a single model and controls its pricing, the value proposition of model-agnostic tooling gets very concrete very fast.

This is not just Microsoft’s play. It is the industry pattern for 2026. Every AI coding tool vendor is converging on the same strategy: proprietary model, usage-based billing, workflow lock-in. The vendors that let you swap models freely are the exception, and they are the ones worth watching precisely because this trend makes them more valuable.

The Take

Microsoft executed a textbook vertical integration in 48 hours, and most Copilot users will not realize it happened until the August bill arrives. The Polaris announcement sounds like a product improvement — faster completions, better benchmark scores, purpose-built for coding. And maybe it is. But the timing reveals the strategy: you do not announce a free model upgrade one day after switching to usage-based billing unless the model change is part of the billing story.

The three-month GPT-4 Turbo fallback is a grace period, not an exit ramp. It buys time for teams to test. It does not buy an alternative. If Polaris degrades your workflow quality — and with any model switch, some workflows will degrade — your options are: accept the regression, retune your prompts for the new model, or migrate off Copilot entirely. None of those are fast.

If your team is on Copilot Business or Enterprise, August is the real deadline. Not to panic, but to test. Get Polaris preview access, run your critical workflows against it, measure the delta. The teams that validate now keep their options. The teams that wait get whatever Microsoft ships.

For everyone else watching the AI coding tool market: this is what the next 12 months look like. Every vendor is building toward owned models with metered billing. The flat-rate era is over. Your tool decisions now need to account for not just what the tool can do today, but who controls the model behind it and what they will charge you tomorrow.

Related

• The Flat-Rate AI Subscription Era Is Ending — the broader pricing shift that set the stage for this move
• GitHub Copilot Tool Profile — full review of Copilot’s capabilities, pricing, and competitive position
• GitHub Copilot SDK Public Preview — context on the SDK layer that makes this migration particularly risky for automation teams