September 22, 2026 · Models & Pricing
SpaceXAI released Grok 4.7 on September 21, pricing the model at $2 per million input tokens and $6 per million output tokens. A faster variant is listed at twice the price for roughly twice the speed, positioning the drop as a bid for coding and knowledge-work workloads rather than a flagship consumer chatbot refresh alone.
Independent roundups flagged a 71% score on DeepSWE among the early benchmark notes. Practitioners have been told to read the spread carefully: headline coding numbers can look strong while general-reasoning gaps remain, and price-per-token is only part of the total cost of an agentic loop.
The launch arrives in a week already crowded with open-weight releases from Chinese labs and renewed debate over whether frontier labs should slow capability gains. Aggressive API pricing is one way a vertically integrated compute player can press incumbents that still carry heavier inference costs.
For teams already routing work across several models, Grok 4.7 is another reason to re-run evals on coding agents, long-context research tasks, and unit-cost per successful job rather than raw tokens.
Key takeaway Grok 4.7 is a price-and-coding play: cheap tokens and a fast tier matter most if your evals show it holding quality on the jobs you actually run.
Photo: Unsplash (neural-network style AI imagery). Sources: AI Weekly (Sept 22, 2026); AIToolsRecap daily brief; Unite.AI model notes.
