The AI arms race has a new scoreboard, and Elon Musk’s xAI is sitting at the top of it. Grok 4.5 just posted a 91.3% score on VulcanBench, an open-source coding benchmark that went live earlier this month, solving 21 out of 23 multi-file software engineering tasks across five programming languages.
That performance puts it ahead of Anthropic’s Claude Fable 5 and OpenAI’s GPT-5.6 Sol.
What VulcanBench actually measures
VulcanBench, which launched its v3 suite on July 10, 2026, evaluates AI coding agents on real-world, pull-request-style tasks, the kind where you’re modifying multiple files across a codebase, not just spitting out a neat function in isolation.
The benchmark’s 23 tasks span Python, Rust, TypeScript, JavaScript, and at least one additional language. Each evaluation runs inside a Docker sandbox, which means results are reproducible and transparent. Cost reporting is baked into the methodology too, so you can see not just whether a model solved the problem, but how expensive it was to get there.
