Grok 4.7 is out: big coding gains, same $2/$6 price
Live now in the API, Cursor and OpenRouter with four reasoning tiers and a 500K window — Terminal-Bench nearly doubles, but the 200K price cliff bites.

Copy markdown
Same price, bigger brain
Grok 4.7 keeps Grok 4.6's $2/$6-per-million pricing but ships a larger base model, extended RL on multi-hour tasks, and better self-verification. It's live now as grok-4.7 on the xAI API, Grok Build, Cursor, OpenRouter and Vercel.
Agentic coding takes the jump
The gains concentrate in long-horizon coding: Terminal-Bench 4.0 climbs to 38.0% at xhigh from 20.3%, DeepSWE v1.1 hits 71.0%, and CursorBench 4.0 reaches 46.3%. If you run coding agents, xhigh is where the delta lives.
Four effort dials, not one
Reasoning is now selectable per call — low, medium, high (default) and xhigh — so you trade latency and tokens for depth. Most of the benchmark lift only shows up at xhigh, so reserve it for the hard tasks.
Mind the 200K price cliff
Under 200K tokens it's $2 in / $6 out ($0.50 cached). Cross 200K and the entire request re-rates to $4 / $12 — not just the overflow. Set a prompt_cache_key to pin cache hits instead of paying full input every turn.
500K window, full tool belt
The 500K-token context carries over, with function calling, web search, X search and code execution built in, plus native Grok Bot harness support for agent loops. A 2x-speed variant exists but is gated to Cursor and Grok Build.