Project Polaris: Microsoft Replaces GPT-4 in GitHub Copilot
Microsoft's in-house AI coding model is ending GitHub Copilot's dependence on OpenAI. Here is what every developer -- free user or paid subscriber -- needs to know before August.
By Free AI News Editorial · · · 9 min read
If you use GitHub Copilot -- whether on the free tier or a paid Pro plan -- the AI model powering your completions is about to change permanently. At Microsoft Build 2026 in San Francisco today, Microsoft unveiled Project Polaris, an in-house coding model years in the making that signals a decisive break from the company's seven-year reliance on OpenAI. Starting in August, every Copilot subscriber will be migrated automatically. This is not a minor update: it is the most significant backend change Copilot has seen since launch, and it arrives at the same time as tightening usage limits and paused sign-ups that are already affecting how developers access the tool for free.
This guide covers what Project Polaris is, how it compares to GPT-4 Turbo, what the August migration means for different Copilot plan holders, and what your best free alternatives look like if the changes push you toward other tools.
What Is Project Polaris and Why Is Microsoft Building Its Own AI Model?
Project Polaris is a coding-specialized large language model built by Microsoft Research and trained primarily on public and private code repositories with an emphasis on security-vetted, enterprise-grade codebases. Unlike general-purpose models such as GPT-4 Turbo, Polaris uses a mixture-of-experts (MoE) architecture where the system routes each inference request to specialized sub-modules tuned for distinct programming languages and frameworks. A query about Rust gets handled by a different sub-module than one about TypeScript or SQL, producing outputs the company claims are meaningfully more accurate than a single dense model handling everything.
The strategic reason for building Polaris is just as important as the technical one. Microsoft and OpenAI ended their exclusive partnership in April 2026, and Copilot had become Microsoft's highest-profile developer product built on top of a competitor's infrastructure. Controlling the model means controlling the cost curve, the latency, the compliance posture, and the roadmap. Polaris runs on Microsoft's custom Maia AI accelerators inside Azure, which Microsoft says reduces per-inference latency and lowers operational cost compared to renting OpenAI capacity. For enterprise customers -- Microsoft's biggest revenue source -- that also means data never leaves Azure's governance perimeter.
Satya Nadella had signaled this direction at Ignite 2025, describing "sovereign AI capabilities across the stack." Project Polaris is that vision arriving on schedule. Azure AI Foundry, Microsoft's unified model development platform, serves as the backbone for Polaris and the family of smaller domain-specific models (SQL generation, security auditing) that shipped alongside it at Build. According to TechTimes reporting from Build today, the SDK now ships a reasoning layer built and operated entirely within Microsoft's stack -- no OpenAI API calls required.
How Does Project Polaris Perform Compared to GPT-4 Turbo for Coding?
Microsoft's benchmarks claim Polaris outperforms GPT-4 Turbo on the two most widely cited coding evaluation suites: HumanEval (function completion from docstrings) and MBPP (mostly basic Python programming problems). The gains are largest in lower-resource languages that GPT-4 Turbo's general training data does not cover as deeply -- Rust and Haskell are the examples Microsoft highlighted. For Python, JavaScript, and Java -- where GPT-4 Turbo has extensive training coverage -- the claimed improvements are narrower, though Microsoft says reduced latency is consistent across all languages due to the Maia hardware advantage.
Here is how the two models break down across key Copilot feature dimensions:
| Feature | GPT-4 Turbo (current) | Project Polaris August 2026 |
|---|---|---|
| Architecture | Dense transformer | Mixture-of-Experts (language-specialized) |
| HumanEval benchmark | Baseline | Higher (margins not fully disclosed) |
| Low-resource language gains | General coverage | Notable gains in Rust, Haskell |
| Pro multi-file context | Standard context window | Up to 100,000 lines |
| Autonomous test generation | Manual trigger | Available in Pro tier |
| Inference infrastructure | OpenAI data centers | Microsoft Maia accelerators (Azure) |
Independent third-party verification of these benchmarks is not yet available -- Microsoft's numbers come from their own evaluation runs. The developer community will need a few weeks of real-world usage to validate the claimed improvements. For teams building on the Copilot SDK, the practical upside is that the same manifest definitions and integrations will work with Polaris without re-architecture, according to Microsoft's Build documentation.
What Does the August 2026 Migration Mean for Current Copilot Users?
The migration is automatic. If you are a Copilot subscriber today, your default model will switch from GPT-4 Turbo to Project Polaris in August without any action required on your part. Microsoft is offering a three-month optional fallback period -- teams that need additional time to test compatibility, evaluate output quality on their specific codebase, or complete ongoing projects can temporarily opt back to GPT-4 Turbo before the full cutover becomes permanent.
The impact varies significantly by which Copilot tier you are on:
- Copilot Free -- The free tier gives access to Copilot in VS Code with monthly usage limits. Polaris replaces the underlying model, but the tier structure and limits remain as currently set.
- Copilot Pro ($10/month) -- Polaris becomes the default. New features like 100,000-line multi-file context and autonomous test generation arrive with Pro tier access. Opus models were already removed from Pro in April.
- Copilot Pro+ ($19/month) -- Full Polaris access plus higher usage limits (Pro+ already offers more than 5x the usage of Pro). Opus 4.7 remains available here, though Opus 4.5 and 4.6 are being retired.
- Copilot for Business / Enterprise -- Polaris rolls out on the same August timeline, with Azure data governance ensuring prompts and completions stay within the enterprise tenant.
The most important action for current subscribers before August: if your team has built workflows that depend on GPT-4 Turbo's specific behavior -- few-shot prompt patterns, output format assumptions, or chain-of-thought structures tuned to GPT-4 -- test those workflows against Polaris now. Once the fallback window closes, there is no guaranteed path back to GPT-4 within the standard Copilot subscription. You can track the AI free tier tracker here for ongoing updates as migration details are confirmed.
Did GitHub Already Tighten Copilot's Free and Individual Tier Limits?
Yes -- and the cuts predated today's Build announcement. GitHub made significant changes to Copilot's individual tiers in late April, affecting both access and model availability in ways that are already live. According to the official GitHub community announcement from April 24, 2026, three changes took effect simultaneously:
- New sign-ups paused: New users cannot currently sign up for Copilot Pro, Pro+, or the Student plan. GitHub framed this as protecting existing customers from degraded service during high-demand periods. There is no stated timeline for when sign-ups will reopen.
- Usage limits tightened: Monthly completions and chat message limits were reduced for Pro plan holders. Usage limits are now visible in VS Code and Copilot CLI so users can monitor their remaining quota. Pro+ subscribers, by contrast, received over 5x the limits of the Pro plan, further widening the gap between the two tiers.
- Opus models removed from Pro: Anthropic's Opus 4.7 is no longer accessible in the standard Pro plan. It is restricted to Pro+ only. Older Opus versions (4.5 and 4.6) are also being retired from Pro+. This is part of a broader model-access differentiation strategy that ensures premium subscribers have exclusive access to the most capable third-party models while standard users get the core Copilot experience.
Read together, April's changes and August's Polaris migration tell a consistent story: GitHub is consolidating its AI supply chain around Microsoft-owned infrastructure, reducing dependency on external model providers, and pushing power users toward higher-margin tiers. The free tier and base Pro plan are being deliberately narrowed so that the most capable features -- Polaris's extended context, autonomous test generation, Opus access -- live exclusively in the more expensive plans.
If you are a student who relied on the free Student plan or an individual developer who signed up recently, check your account status now. The pause on new sign-ups means you cannot currently access Copilot's paid tiers even if you want to upgrade. See the full breakdown of Copilot's billing changes for more context on how usage-based charges stack alongside these limits.
What Are the Best Free Alternatives to GitHub Copilot Right Now?
If the tightening limits or paused sign-ups push you to look elsewhere, the free AI coding landscape is stronger than it has ever been. Several tools have expanded their free tiers precisely as GitHub has contracted its own:
- Cursor (Free tier) -- 2,000 completions per month, GPT-4o and Claude Sonnet 4 available. The free tier is competitive for individual developers not needing enterprise features. Cursor's agent mode, which can edit multiple files and run terminal commands, is available without a paid plan. We covered the Cursor vs Windsurf vs Zed free tier breakdown in detail.
- Codeium / Windsurf (Free) -- Unlimited completions in their free tier with no monthly cap, alongside a generous chat quota. Windsurf's Cascade agent mode is one of the strongest free agentic coding experiences currently available.
- Zed with AI (Free) -- Open-source editor with native AI assistant integration. Supports multiple model backends including local models, making it useful for privacy-sensitive projects or offline environments.
- Amazon CodeWhisperer (Free tier) -- Free for individual developers, with no usage cap on code suggestions and 50 security scans per month. A strong pick for developers already in the AWS ecosystem.
- Continue.dev (Open Source) -- A fully open-source Copilot alternative that works in VS Code and JetBrains. Bring your own API keys or connect to a local Ollama instance for zero ongoing cost.
For developers who want to self-host their entire coding AI stack, OpenCode and other open-source coding agents now offer capabilities that rival hosted tools. The barrier to self-hosting has dropped significantly in 2026, and for teams with a spare GPU it is a credible path to escaping usage limits entirely.
🔑 Key Takeaways
- Project Polaris, announced at Microsoft Build 2026 today, will automatically replace GPT-4 Turbo as the default Copilot model for all subscribers in August 2026, with a three-month fallback option for teams needing extra transition time.
- Polaris uses a mixture-of-experts architecture with language-specific sub-modules and runs on Microsoft's Maia AI accelerators, which Microsoft says reduces latency and cost while keeping data within Azure's governance perimeter.
- GitHub has already tightened Copilot's individual tiers: new sign-ups for Pro, Pro+, and Student plans are paused, usage limits have been cut, and Opus models have been removed from the standard Pro plan as of April 2026.
- Pro subscribers gain new Polaris-specific features including multi-file context up to 100,000 lines and autonomous test generation, but only if they are already on a paid plan -- the paused sign-up window blocks new joiners.
- Free alternatives including Cursor, Windsurf, Codeium, and open-source tools like Continue.dev have all maintained or expanded their free tiers, making this a realistic moment to evaluate Copilot alternatives if the changes do not work for your workflow.
Frequently Asked Questions
What is Project Polaris in GitHub Copilot?
Project Polaris is Microsoft's in-house AI coding model that will replace GPT-4 Turbo as the default engine in GitHub Copilot starting August 2026. It uses a mixture-of-experts architecture with language-specialized sub-modules, outperforms GPT-4 Turbo on HumanEval and MBPP benchmarks, and runs on Microsoft's custom Maia AI accelerators inside Azure.
When will Project Polaris replace GPT-4 in GitHub Copilot?
The automatic migration starts in August 2026. All Copilot subscribers will be moved to Polaris as their default model. Microsoft is providing a three-month optional fallback so teams can stay on GPT-4 Turbo temporarily if they need to test compatibility before the permanent cutover.
Is GitHub Copilot free still available in 2026?
A limited free tier exists, but GitHub paused new sign-ups for Copilot Pro, Pro+, and Student plans in April 2026. Existing subscribers are unaffected, but new users cannot currently join the paid individual tiers. The free Copilot tier within VS Code remains available with monthly usage caps that have been tightened as of the April update.
Does Project Polaris outperform GPT-4 Turbo for coding?
Microsoft claims Polaris scores higher than GPT-4 Turbo on HumanEval and MBPP coding benchmarks, with the biggest gains in lower-resource languages like Rust and Haskell where Polaris's specialized sub-modules provide an advantage. Independent verification of these benchmarks is pending. Latency improvements are consistent across all languages due to the Maia accelerator hardware.
What happened to Opus models in GitHub Copilot?
As of April 24, 2026, Anthropic's Opus models were removed from Copilot Pro plans. Opus 4.7 remains available exclusively in Copilot Pro+ plans. Opus 4.5 and 4.6 are being retired from Pro+ as well. The changes are part of a broader model-access tightening that restricts the most capable third-party models to higher-tier subscribers only.