Grok 4.7 codes better at the same price: vendor claims nearly twofold growth on Terminal-Bench

SpaceXAI released Grok 4.7 on 21.09.2026 without touching pricing: $2 input, $6 output per million tokens, and $0,50 for cached input. The upgrade goes to everyone paying by the token, with no budget recalculation.

The vendor reported gains. On CursorBench 4.0, Grok 4.7 scores 46,3% versus 40,4% for Grok 4.6. On Terminal-Bench 4.0, where the model operates in a terminal on its own, the score rose from 20,3% to 38,0%. DeepSWE v1.1 increased from 65,2% to 71,0%, while EEBench for electrical engineering rose from 53,0% to 64,0%.

Third-party runs paint a more modest picture. Artificial Analysis gave Grok 4.7 46 points in its v4.3.2 index, while Claude Fable 5.1 and GPT-6 each score 53. An independent Terminal-Bench 4.0 run showed 26% for the new model instead of the vendor's claimed 38%, while GPT-6 Astra scored 60%.

Where to get it. Grok 4.7 is already available in Cursor and Grok Build, through the API under the `grok-4.7` identifier with a key from console.x.ai, as well as in third-party coding tools, model routers, and clouds. The CLI installs with one line: `curl -fsSL https://x.ai/cli/install.sh | bash`.

The context window is 500 thousand tokens, and the model's knowledge cuts off in May 2026. A prompt longer than 200 thousand tokens doubles the rate for the entire request, to $4 for input and $12 for output. There is a fast variant: it responds twice as fast, with output priced at $12 per million tokens.

While third-party benchmarks keep Grok 4.7 below GPT-6 and Claude Fable 5.1, price remains its strongest argument.