Grok 4.6 is exceptional at fixing React code It ranks #2 on ReactBench:

Aiden Bai's post backs its claim with an 11-second clip rather than a link. The chart in that clip, titled ReactBench Pass@1, puts xAI's Grok 4.6 at xhigh effort on 38%, second only to GPT 5.6 Sol at max effort on 48% and ahead of Grok 4.5, GPT 5.6 Terra, Opus 5 and Fable 5. ReactBench is not an outside leaderboard: its README describes it as a product of Million Software, the company linked from Bai's profile, and asks interested teams to email aiden@million.dev.

The benchmark exists because, in its authors' words, models can pass every test and still write React that fails in production. Each task comes from real open-source React code. Writing React tasks ask the agent to implement a change taken from a merged pull request; Fixing React tasks ask it to remove specific problems flagged by React Doctor, an open-source code scanner. A solution scores only if the task's required tests pass (hidden behaviour tests, in the writing tasks) and it introduces no new graded React Doctor issue. The methodology page lists 39 tasks in the current release, 28 for writing and 11 for fixing.

The clip shows a single pass@1 chart, not a separate fixing score, and the board has moved since August. On the current v1.2 leaderboard, which now includes models such as Opus 5.5 and Fable 5.1, Grok 4.6 scores about 34% and is tied for ninth. On the Fixing React subset it scores 18.2% at xhigh effort, sixth on that table, behind GPT 5.6 Sol at 27.3%, Apex at 21.8% and three models at 20%. The #2 ranking describes the earlier chart; it is not where Grok 4.6 stands on the benchmark today.