What Is Microsoft Project Polaris? Copilot’s New Model

Microsoft built its own coding model to power GitHub Copilot. Here is what Project Polaris actually is, why it exists, and what changes for subscribers in August 2026.

What is Microsoft Project Polaris?

Microsoft Project Polaris is the company’s first in-house AI coding model, built to replace OpenAI’s GPT-4 Turbo as the default reasoning engine inside GitHub Copilot. Microsoft revealed it at Build 2026 on June 2, 2026, and confirmed it will become the default model for all Copilot subscribers starting in August 2026. In plain terms: the brain behind the autocomplete, chat, and agent features that millions of developers use every day is being swapped out for one Microsoft trained, owns, and runs itself.

Architecturally, Microsoft describes Project Polaris as a mixture-of-experts (MoE) model with specialized sub-modules tuned for different programming languages and frameworks. Rather than a single monolithic network handling every request, an MoE design routes each prompt to the experts most relevant to the task, which can lower the compute cost per token while keeping a large total parameter count. Microsoft says the model also uses inference-time reasoning techniques to handle complex, multi-file refactoring rather than just single-function completion.

It is important to be precise about status. As of early June 2026, Project Polaris is ANNOUNCED, not yet shipping to users. It is scheduled to become generally available as the Copilot default in August 2026. Microsoft has not published a public preview build, a model card, a parameter count, or a hard benchmark table. Everything known so far comes from the Build 2026 keynote and press coverage, so treat performance claims as vendor statements until independent results appear.

Why did Microsoft build its own coding model?

Microsoft built Project Polaris to own its most-used developer product end to end and to stop depending on OpenAI for the model layer of GitHub Copilot. The timing is not a coincidence. On April 27, 2026, Microsoft and OpenAI announced the next phase of their partnership, formally ending the seven-year exclusive arrangement. Under the new terms, Microsoft’s license to OpenAI’s IP became non-exclusive (and runs through 2032), OpenAI gained the freedom to sell its products on any cloud, and Microsoft stopped paying a revenue share on OpenAI products it resells. Project Polaris is what ‘reduced OpenAI dependence’ looks like in practice.

There are three concrete motivations beyond strategy theater. First, cost: Microsoft says Polaris runs on its custom Maia AI accelerators inside Azure, which it claims lowers per-inference latency and operational cost versus the GPT-4 backend. When you serve a coding assistant to a paying base reported at 4.7 million subscribers and 77,000 enterprise customers, even a small per-token saving compounds into real margin.

Second, control. Owning the model means Microsoft can tune it specifically for Copilot’s workloads, ship updates on its own schedule, and avoid the commercially awkward situation of paying a partner whose ChatGPT and Codex products compete for the same developers. Third, this is part of a wider MAI (Microsoft AI) push: at Build 2026 the company positioned Polaris alongside a broader ‘MAI v2’ effort to replace OpenAI-supplied models across its product surface, including newly commercialized speech, voice, and image models.

How does Project Polaris compare to GPT-4 Turbo?

Microsoft claims Project Polaris outperforms GPT-4 Turbo on the HumanEval and MBPP coding benchmarks, with the biggest gains in low-resource languages like Rust and Haskell, but it has not released any specific scores and the figures are not independently verified. That caveat matters more than the claim. HumanEval and MBPP are old, saturated benchmarks where frontier models already cluster near the ceiling, so ‘beats GPT-4 Turbo on HumanEval’ is close to meaningless as a 2026 differentiator. For a real comparison you would want SWE-bench Verified or a similar agentic benchmark, and Microsoft has published neither.

What we can compare is the shape of the two offerings rather than raw quality. The table below lays out the verified differences between the outgoing default and the incoming one. Note that several Polaris-tier features (the 100,000-line multi-file context and autonomous test generation) are described as Pro-tier capabilities, and pricing for Polaris was not disclosed at Build.

The honest takeaway: this is not framed as a leap to a smarter model. It is framed as a comparable-or-better model that Microsoft owns and can run more cheaply. If Polaris merely matches GPT-4 Turbo while cutting Microsoft’s cost and removing a dependency, that is a win for Microsoft regardless of whether developers feel a quality difference.

What changes for GitHub Copilot users in August 2026?

Starting in August 2026, GitHub Copilot will automatically migrate subscribers from GPT-4 Turbo to Project Polaris as the default model, with an optional three-month fallback to GPT-4 Turbo that you must configure before the switch. For most users this should be a behind-the-scenes change: the Copilot UI, the keybindings, and the workflow stay the same. The model answering your prompts is what differs. That is precisely why Microsoft can do it at this scale without a disruptive rollout.

The practical to-do list is short but real. If your team relies on specific model behavior, deterministic outputs, or has tuned prompts and CI checks around GPT-4 Turbo’s quirks, set up the fallback before August and validate Polaris against your own test suite during the overlap window. Pro-tier subscribers stand to gain the most net-new capability, since the 100,000-line multi-file context and autonomous test generation are pitched as Pro features.

Two open questions remain unanswered as of early June 2026. First, pricing: Microsoft disclosed no price change for Polaris, so assume current Copilot pricing holds until told otherwise. Second, what happens to personal-plan users after the three-month fallback expires. Coverage confirms the enterprise and team fallback mechanism but is silent on long-term opt-out for individual plans, so watch GitHub’s official changelog for specifics.

How does Project Polaris fit Microsoft’s OpenAI-independence push?

Project Polaris is the clearest signal yet that Microsoft is moving from reselling OpenAI’s models to running its own, completing a strategy that the April 2026 partnership reset made possible. The renegotiated deal did not break the relationship: Microsoft remains OpenAI’s primary cloud partner, OpenAI products still ship first on Azure, and the IP license runs through 2032. What changed is optionality. Microsoft is now free to build, ship, and default to its own models in its own products, and Copilot is the highest-profile place to do it.

Zoom out and Polaris is one piece of a multi-model posture. Microsoft is repositioning both GitHub Copilot and Azure AI as platforms that route across OpenAI, Anthropic’s Claude, open-source models, and now Microsoft’s own MAI family. Microsoft Foundry, the company’s model catalog, now lists thousands of models. The message to enterprises is that the underlying engine is swappable and that orchestration, not any single provider, is where the durable value sits.

For OpenAI, losing the Copilot default is a meaningful symbolic and commercial hit, even if it keeps the cloud relationship. For the broader market, this is the commoditization thesis made concrete: when a platform owner can replace the model under millions of paying users without changing the product, the model layer starts to behave like an interchangeable component rather than the product itself.

What else did GitHub Copilot announce at Build 2026?

Alongside Project Polaris, Build 2026 brought multi-agent support to VS Code and pushed Copilot Workspace to general availability, signaling Microsoft’s shift from single-agent assistance to orchestrated, multi-agent development. The multi-agent VS Code extension uses an orchestrator agent that decomposes an objective and spawns parallel subagents assigned to discrete workstreams such as linting, test generation, documentation, and security review. Developers can watch progress in real time and steer mid-run. It became available for adoption at Build on June 2, 2026, and is positioned as a step beyond the single-agent mode that reached general availability in March 2026.

Copilot Workspace, meanwhile, exited beta and hit general availability at Build, described by GitHub’s CEO as the biggest change to Copilot since launch. The GA release adds a Fleet mode for autonomous narrow-scope operations and an Autopilot mode for scheduled background task execution. A more aggressive Autonomous Agent Mode for enterprise, which can write, test, and commit entire feature branches, is slated for July 2026 and remains in preview.

Taken together, the Build 2026 Copilot story is two-pronged: Polaris changes who supplies the intelligence, while the multi-agent and Workspace features change how that intelligence is applied. The model swap is the supply-chain move; the agent features are the product move. Both point at the same destination, a Copilot that does more on its own with Microsoft owning every layer underneath it.

Should you trust Microsoft’s Project Polaris benchmark claims?

Be skeptical of the benchmark claims until independent results land, because Microsoft has only said Polaris beats GPT-4 Turbo on saturated benchmarks and has released no specific numbers. Press coverage is explicit that the figures are Microsoft’s own and have not been confirmed by independent auditors. HumanEval and MBPP, the two benchmarks Microsoft cited, are widely regarded in 2026 as saturated, with frontier models bunched near the top. A win on those tells you the model is competent, not that it is better than the alternative you are using today.

What would change the picture is a result on a discriminating, agentic benchmark, paired with a model card and transparency on training data. None of that exists yet. The most credible signal Microsoft offered is indirect: it is willing to put Polaris in front of millions of paying developers as the default. A company does not do that with a model it believes is clearly worse, even if ‘clearly worse’ and ‘clearly better’ are both unproven from the outside.

The pragmatic stance for developers and engineering leaders is to run your own evaluation. During the fallback window, point Polaris and GPT-4 Turbo at your real tasks, your real codebase, and your real acceptance criteria. Vendor benchmarks are a starting hypothesis; your task suite is the only test that matters for your team. Until then, the safe summary is that Polaris is a strategically important, technically plausible, but externally unverified model.

Builder’s take

I ship two AI products, and I route across model providers every day. Microsoft Project Polaris is less a model story than a supply-chain story. Here is what I would watch if I depended on Copilot.

• The headline is not the benchmark, it is ownership. Microsoft now controls the model, the Maia silicon, and the IDE. That vertical stack is the real moat, not a HumanEval delta.
• Microsoft did not publish a single hard number. ‘Outperforms GPT-4 Turbo on HumanEval and MBPP’ is a saturated-benchmark claim that tells you almost nothing in 2026. Treat it as marketing until SWE-bench Verified results land.
• The three-month fallback window is the only consumer-protection detail that matters, and you have to opt in before August. If your CI depends on model behavior, configure it now and benchmark Polaris against your own task suite, not Microsoft’s.
• This is the commoditization thesis playing out in public: when the platform owner swaps the engine under 4.7 million paying users without changing the UI, the model layer is officially a component, not the product.

Frequently asked questions

Primary sources

• GitHub Copilot Replaces GPT-4 With Project Polaris, Ships Multi-Agent VS Code at Build — Tech Times
• Microsoft Build 2026 Recap: Project Polaris Cuts the OpenAI Cord — ChatForest
• Microsoft Just Built Its Own AI Model to Replace OpenAI’s Inside GitHub Copilot — FourWeekMBA
• The next phase of the Microsoft-OpenAI partnership — Microsoft (Official Blog)
• OpenAI shakes up partnership with Microsoft, capping revenue share payments — CNBC
• Microsoft Build 2026: Homegrown AI Models to Power GitHub Copilot — Windows News

Last updated: June 2, 2026. Related: Products.