Microsoft says MAI Code 2 and Code 2 Flash were reinforcement-trained directly in production Copilot CLI and VS Code agent harnesses. Its MSBench comparisons use matching CLI settings, maximum reasoning, no subagents and three independent runs averaged as a single-attempt pass rate; infrastructure failures count as failures.
The smaller Flash leads SWE Atlas, 44.8% to 43.1%, while Code 2 leads DeepSWE, 73.2% to 59.9%. The former mixes codebase questions, tests and refactoring; the latter requires hidden tests to pass.
Microsoft’s four-test average gives Code 2 63.2% at about $0.31 per task, against Sol’s 64.1% at $0.76 and Opus’s 65.3% at $6.31. Code 2’s cost includes a 50% promotion. Three tests use Microsoft’s harness, while FrontierCode figures come from Artificial Analysis. Flash’s advertised 61% and roughly $0.05 use only three tests, also at promotional pricing. Task-level comparisons are more revealing than a single ranking.