xAI has launched Grok 4.7 with the same headline API price as Grok 4.6, while early benchmark results suggest a sharp split between cheap coding assistance and dependable autonomous work. For developers, the useful question is not whether Grok 4.7 is “frontier” in the abstract, but whether its lower bill survives contact with the task they actually need done.
xAI Watch analysis
What happened
Grok 4.7 shipped on 21 September through the Grok API, grok.com and the iOS and Android apps, with distribution through developer tools including Cursor. xAI positions it as a coding and knowledge-work model rather than a general chat refresh.
The standard API tier costs $2 per million input tokens and $6 per million output tokens for prompts below 200,000 tokens. Above that threshold, prices double to $4 and $12 respectively. The model has a 500,000-token context window, while a faster variant costs twice the standard rate and is aimed at latency-sensitive coding agents.
What the results show
- A genuine price advantage
Grok 4.7 undercuts the headline token prices of comparable OpenAI and Anthropic offerings by a considerable margin. - Mixed coding performance
It scored 46.3% on CursorBench 4.0 and 71.0% on DeepSWE v1.1, close to or ahead of some rivals on those tests. - A serious agentic weakness
On Terminal-Bench 4.0, Grok 4.7 scored 26%, compared with 60% for GPT-6 Astra and 55% for Claude Fable 5.1. - A costly context cliff
Crossing 200,000 tokens doubles input and output rates, so long codebases and document chains need careful budgeting.
Why it matters
The numbers describe two different products hiding under one model name. Grok 4.7 may be attractive for high-volume work where a person remains closely involved, especially when price matters more than squeezing out the last few points of accuracy.
The Terminal-Bench result is more troubling for teams hoping to hand an agent a messy repository and let it work through many commands without supervision. A cheap model that drops the thread can become rather expensive once a person has to rescue it.
Our read
Grok 4.7 is a credible price-war launch, not a clean victory lap. The sensible approach is workload testing: use the cheaper model where its strengths fit, but do not confuse a low token price with low total cost. The spreadsheet has entered the benchmark arena, and it has brought receipts.
What to watch
- Whether xAI improves long-horizon tool use in later updates.
- Whether independent teams reproduce the large Terminal-Bench gap.
- How the 200,000-token pricing threshold affects real production bills.
- Whether Cursor and other integrations expose useful routing or fallback controls.
Discussion spark: Would you choose Grok 4.7 for its much lower price despite the weak long-horizon agent score, or is reliability on multi-step work worth paying substantially more for?
Sources and evidence
- Grok 4.7 Pricing vs GPT-6 Astra, Claude Fable 5.1 – shattered.io (22 September 2026, 01:01 UTC)
not affiliated with or endorsed by xAI, SpaceXAI, SpaceX, X Corp. or Cursor