Exa open-sources WebCode coding benchmark suite

This is a useful benchmark because it measures the failure mode that actually derails agents: noisy retrieval, not just LLM reasoning. It also doubles as a strong product narrative for Exa's search API.

• –WebCode splits evaluation into contents quality, highlights, retrieval quality, and sandboxed coding tasks, so teams can pinpoint whether failures come from extraction, ranking, or downstream reasoning.
• –The groundedness metric is the sharpest idea here: it distinguishes "the model answered correctly" from "the search system actually surfaced the right evidence."
• –Exa reports 82.8 completeness, 94.5 signal, and 96.7 code recall on its contents track, ahead of Parallel and Claude on several axes, which gives the suite real bite beyond the announcement post.
• –The released datasets are practical, not toy examples: 250 URLs, 317 QA pairs, and coding tasks seeded from recent library releases with hidden discriminators make the suite feel closer to production agent work.
• –For coding-agent builders, this is more actionable than a generic leaderboard because it tests whether retrieved context is precise enough to drive code changes.