Grok 4.7

Grok 4.7, released September 21, 2026, was the model Elon Musk had trailed for most of the month: a larger successor to Grok 4.6 that he said traded some serving speed for capability. Its model card, dated the same day, is more specific than the pitch was — supplemental training ran longer than Grok 4.6's on curated reasoning data and engineering corpora, agentic reinforcement learning covered kernel optimisation, web development and computer-aided design, and the model took supplemental training on anonymised Cursor workflow data, the clearest use yet of the coding company SpaceXAI had bought earlier in the year. Pretraining data stopped in June 2026, with supplemental data as late as August. It shipped with a 500,000-token context window and at Grok 4.6's prices, unchanged in every column — $2 per million input tokens and $6 per million output under 200K, doubling to $4 and $12 above it — and went out through the API, the Grok Build terminal agent, Cursor on every plan tier, the Office add-ins and gateways including OpenRouter, with the consumer apps to follow later. The card gave no parameter count, leaving the 2.1 trillion Musk had claimed before launch unconfirmed.

Engineering was where the card put its weight, and the picture was one of a strong second place. Grok 4.7 scored 66.0% on atopile's EEBench for circuit and chip design, ahead of Grok 4.6 at 60.0% and Claude Opus 5 at 61.6% but behind GPT-6 Astra's 69.3% — and xAI's own launch chart had put the figure at 64.0%, a discrepancy the card did not explain. It took CADGenBench outright at 44.4%, and the legal row at 19.6% on Harvey's Legal Agent Benchmark, against 15.8% for Grok 4.6 and 2.5% for GPT-5.6 Sol. Knowledge work came second to Claude Fable 5.1 on both boards, at a GDPval-AA v2.1 Elo of 1695 and 1657 on AA-Briefcase v1.1. The coding results were mixed in a way the launch post was not: 71.0% on DeepSWE v1.1 and 46.0% on SWE-Marathon v1.1 sat just behind GPT-5.6 Sol and Claude Opus 5, 38.0% on Terminal-Bench 4.0 trailed Claude Fable 5.1's 57.9%, and on FrontierSWE V2 — thirty-four challenges with twenty hours each — its 29.0% was barely half of Fable 5.1's 56.3%. Elsewhere it reported 56.7% on HealthBench Professional, 44.5% on LatchBio's biology suite, and small cyber gains it framed as defensive: 80.3% on CyberGym and 29% on the hard subset of the third-party CathedralBench, against 25% for Grok 4.6. On the dual-use biology probes the card reported no gain at all and in several places a decline, which it attributed to safer training environments rather than to heavier refusal training.