A cheaper Grok 4.7 token does not make the task cheaper: usage doubled versus 4.6

Artificial Analysis's measurement from 21.09 turns xAI's promise into an actual bill. The per-token rate stayed the same, but tokens per task doubled, so the same work costs more than it did on the previous model.

The price did not move. The API charges $2 per million input tokens, $6 for output, and $0.50 for cached input. For requests with contexts of 200 thousand tokens or more, all three rates double. The fast variant is sold separately at $12 per million output tokens.

xAI does not measure its claim of “twice as fast and twice as cheap” against its predecessor. The company explicitly says that 4.7 is offered at the same price and speed as 4.6, and is twice as cheap as comparable competitor models.

The measurements diverge. In xAI's own runs, the new model beats 4.6 on every benchmark; on DeepSWE v1.1, it scores 71.0% versus 65.2%. Artificial Analysis gives it 46 points on its Intelligence Index, two above 4.6 and below Claude Fable 5.1 and GPT-6, both at 53.

The gap is wider on agentic Terminal-Bench 4.0. Grok 4.7 scores 26%, while Claude Fable 5.1 reaches 55% and GPT-6 Astra 60%.

An average task takes 7.1 minutes, with output at around 188 tokens per second. The hallucination rate fell to 29% from 34%.

In the API, the model is called grok-4.7, the key is stored in the XAI_API_KEY variable, and calls go through the xAI SDK, OpenAI SDK, or curl. The context window is 500,000 tokens. In Cursor, simply select “Grok 4.7” from the model list.

Alongside the API and Cursor, the model is available through Grok Build, OpenRouter, Vercel, and Cloudflare. xAI's docs recommend setting prompt_cache_key, so repeated context is billed at $0.50 per million.

In its own Grok Build harness, the new model scores 56 points on the Coding Agent Index versus 47 for the 4.6-based build, placing 4th among native harnesses.