Mistral Vibe for Code Leads Coding Agent Benchmark

Mistral Vibe for Code achieved the highest score of 22 out of 25 in a comparative benchmark evaluating four leading coding agents on a practical scaffold-to-pull request task. The evaluation, conducted as of July 14, 2026, assessed agents on feature scaffolding, test generation and execution, pull request workflow, surface coverage, and cost/openness/control.

The benchmark tasked agents with adding a new `/subscriptions` endpoint to an existing Python/FastAPI service, including scaffolding routes, Pydantic models, service layers, generating and running tests, fixing failures, and opening a pull request. Mistral Vibe for Code received top marks (5/5) in surface coverage and cost/openness/control, with strong scores of 4/5 in scaffolding, test loop, and PR/async workflow.

Claude Code and OpenAI Codex tied for second place with scores of 21/25. Claude Code excelled in scaffolding, test loop, and PR/async workflow (all 5/5) but scored lower in cost/openness/control (2/5). OpenAI Codex also scored 5/5 in PR/async workflow and surface coverage, with 4/5 in scaffolding and test loop, and 3/5 in cost/openness/control.

Cursor ranked fourth with 16/25, demonstrating weaker performance across most categories, particularly in the test loop (3/5) and PR/async workflow (3/5). The benchmark methodology emphasized that scores reflect documented features, published benchmarks, and vendor specifications, with caveats regarding the direct comparability of different benchmark suites like SWE-bench Verified, SWE-Bench Pro, and Terminal-Bench. Vendor claims, such as Mistral's cost-efficiency figure, were noted as self-reported.