What Microsoft Built—and Why It’s Different

MAI-Code-1-Flash is Microsoft’s first coding model trained entirely in-house, without OpenAI data or third-party model distillation. It shipped on June 2, 2026 at Build 2026 and is rolling out automatically to GitHub Copilot users in Visual Studio Code. If you pay for Copilot, you may already be using it without knowing.

The architecture is a sparse Mixture-of-Experts with 137 billion total parameters, but only 5 billion activate per inference token. That’s the same design principle behind many frontier models: a massive parameter bank for specialization, low active compute for fast, cheap inference. The context window is 256,000 tokens.

What sets it apart is the training environment. Microsoft trained MAI-Code-1-Flash directly on the GitHub Copilot production harness—the actual file-editing tools, terminal integrations, and multi-step task loops that developers use every day. Offline improvements were validated by evaluating checkpoints on real GitHub Copilot usage patterns, from repository Q&A to telemetry-grounded refactoring tasks. The goal was to close the gap between benchmark performance and real Copilot behavior. As the Microsoft AI team put it: “We built MAI-Code-1-Flash with production workflows at the center, rather than optimizing only for benchmarks.”

Head-to-Head Benchmark Numbers

Microsoft ran both models through the same GitHub Copilot production harness, measuring both pass rate and average token usage per task. Here’s what they found:

The headline number is SWE-Bench Pro, where the gap widens to 16 percentage points (51.2% vs 35.2%). SWE-Bench Pro tests on harder, more diverse real-world GitHub issues than the standard Verified set, and the spread here matters—Haiku 4.5 struggles considerably more on the harder task distribution.

Terminal Bench 2, which evaluates models on completing tasks in a terminal environment, shows a similar 13-point gap. And on instruction-following—where models are tested on whether they precisely follow complex, multi-constraint requests—MAI-Code-1-Flash leads by nearly 29 points on IF Bench. Strong instruction-following is what makes an agentic coding assistant actually usable in real developer workflows.

Beyond pass rates, Microsoft also measured token consumption. MAI-Code-1-Flash solved harder problems with up to 60% fewer tokens than Claude Haiku 4.5 on SWE-Bench Verified. Fewer tokens means lower inference cost and lower latency—two things that compound fast when you have hundreds of developers running agentic workflows all day.

The Benchmark Methodology Question

Before concluding that MAI-Code-1-Flash simply outperforms Claude Haiku 4.5, there’s a discrepancy worth flagging. Anthropic’s own published numbers put Claude Haiku 4.5 at 73.3% on SWE-Bench Verified. Microsoft’s testing shows 66.6% for the same model on the same benchmark. That’s a 6.7 percentage point gap—large enough to flip the headline result.

The most likely explanation is the harness. Anthropic benchmarks Haiku 4.5 using their own inference setup—optimal prompting, scaffolding, retry logic, and evaluation conditions tuned for the model. Microsoft ran Haiku 4.5 through the GitHub Copilot production harness, which is what MAI-Code-1-Flash was specifically built and trained for. Haiku 4.5 was not.

This isn’t a gotcha. It’s actually the honest framing for this comparison. A model trained end-to-end in a specific tool environment will outperform a general-purpose model dropped into that same environment—almost by definition. The real question is whether the advantage holds in a neutral, harness-independent evaluation. Microsoft hasn’t published those numbers yet.

Take the SWE-Bench Verified result as a case in point. Anthropic says Haiku 4.5 scores 73.3%. Microsoft says 66.6% in their harness. MAI-Code-1-Flash scores 71.6% in that same harness. Run all three numbers together and MAI-Code-1-Flash doesn’t actually beat Haiku 4.5’s peak score—it beats it in the Copilot environment. For Copilot users, that’s the number that matters. For API developers, Anthropic’s number is the more relevant baseline.

Token Efficiency: Where the Cost Advantage Lives

The 60% token reduction on hard tasks is the claim most worth scrutinizing—and it’s also the one with the most tangible business case. Microsoft achieves this through a mechanism they call “adaptive solution length control.” The model adjusts reasoning depth to the complexity of the task: short, fast responses for simple completions; deeper multi-step reasoning for complex refactoring or cross-file changes.

Claude Haiku 4.5 doesn’t explicitly document an equivalent mechanism, though Anthropic’s extended thinking mode (available on Haiku 4.5) provides some of this capability when enabled. The difference is that with MAI-Code-1-Flash, the adjustment is automatic and baked into the base model behavior, not an opt-in feature that costs extra tokens to activate.

For an enterprise team running thousands of Copilot sessions per day, a 60% token reduction compounds aggressively. Microsoft hasn’t published API pricing for MAI-Code-1-Flash (it currently only ships inside Copilot), but the efficiency story is already meaningful for Copilot unit economics. For context, Claude Haiku 4.5 via Anthropic’s API costs $1.00 per million input tokens and $5.00 per million output tokens—reasonable for API users, but not the relevant comparison for Copilot’s bundled pricing model.

Who Should Use Which Model

These models serve different users in practice, and the choice is often made for you by which ecosystem you’re in.

MAI-Code-1-Flash is the right choice if: you’re a GitHub Copilot user in VS Code and you want the model that was built and optimized for exactly that environment. You don’t need to configure anything—it ships automatically through the model picker and the Auto picker. This also applies to teams that care most about token efficiency and cost-per-task in agentic Copilot workflows.

Claude Haiku 4.5 is the right choice if: you’re accessing AI via the API to build your own tooling, you need extended thinking mode, computer use, or image input (Haiku 4.5 is multi-modal; MAI-Code-1-Flash is text/code only), or you’re not in the GitHub Copilot ecosystem at all. Haiku 4.5 also shines in multi-agent setups where you want sub-agents running in parallel at low cost—Anthropic has built significant infrastructure around this use pattern.

The honest summary: these aren’t direct competitors for most developers right now. Copilot users get MAI-Code-1-Flash automatically. API developers won’t have access to MAI-Code-1-Flash at all. This comparison matters most if you’re a developer evaluating whether to build workflows inside GitHub Copilot versus an API-first stack, or if you’re following Microsoft’s broader push to build its own model family independent of OpenAI.

The numbers show that training a model for a specific production environment produces real gains in that environment. Whether those gains survive a neutral comparison is the question Microsoft’s benchmark setup can’t yet answer. For the roughly 1.8 million paid Copilot users, the production harness numbers are the ones that matter—and on those, MAI-Code-1-Flash leads. For everyone else, the coding assistant landscape is broader than this head-to-head suggests.

Further Reading

• Introducing MAI-Code-1-Flash — Microsoft AI — the full technical blog post with benchmark charts, architecture notes, and the sample apps built with the model
• Claude Haiku 4.5 — Anthropic — Anthropic’s official specs and benchmark numbers, including the 73.3% SWE-Bench Verified figure that differs from Microsoft’s harness result
• MAI-Code-1-Flash: Copilot-Native Coding Model Has Different Benchmarks Than You’d Expect — ChatForest — independent analysis of why the benchmark numbers look different depending on who’s measuring