xAI releases Grok 4.7, a cheaper coding/knowledge model that trails Claude and GPT-6 on independent benchmarks

xAI launched Grok 4.7, priced at $2 per million input tokens and $6 per million output tokens, available through the Grok API, Cursor, and Grok Build. On the Artificial Analysis Intelligence Index (v4.3.2), it scores 46 overall against 53 for both Claude Fable 5.1 and GPT-6. On Terminal-Bench 4.0, a coding-agent benchmark, it scores 26%, behind GPT-6 Astra (60%), Claude Fable 5.1 (55%), and even the cheaper DeepSeek V4.1 Flash (27%).

Anyone choosing a model for coding or agentic tasks now has a concrete price-versus-capability data point: Grok 4.7 is cheap but currently the weakest of the named frontier and near-frontier models on independent coding benchmarks. It is not a reason to switch away from Claude or GPT-6 for serious coding work, but it may be relevant for low-stakes, high-volume tasks where cost matters more than accuracy. For xAI, this is evidence the company has not closed the capability gap despite a bigger base model and longer RL training.

xAI's framing of Grok 4.7 as its "most capable model yet" and emphasis on "improved self-verification" are marketing claims not tested by the cited benchmarks. The benchmark figures come from Artificial Analysis via The Decoder, with no direct links to the xAI announcement or the benchmark pages in the underlying data, so these numbers cannot be independently re-verified from this article alone. Low pricing is being framed as a competitive strength, but a 20-34 point benchmark deficit against similarly priced or cheaper alternatives undercuts that pitch.

1. 01Whether Artificial Analysis or another independent benchmarker publishes updated or disputing scores for Grok 4.7 in the coming weeks
2. 02Adoption signals: whether Cursor, Grok Build, or other coding tools report meaningful usage share shifting to Grok 4.7 given its price
3. 03xAI's next model release timeline and whether it narrows the Terminal-Bench and Intelligence Index gap versus Claude and GPT-6, which would indicate a real capability trajectory rather than a one-off cheap release