SpaceXAI launches Grok 4.7 for coding and long-running knowledge work

SpaceXAI launched Grok 4.7, a new model for coding, agentic tasks and knowledge work, with launch-day access through the SpaceXAI API, Grok Build and Cursor. The release gives developers a 500,000-token context window and standard API pricing starting at $2 per million input tokens and $6 per million output tokens.

On the SpaceXAI API, the model ID is grok-4.7. It accepts text and image inputs, returns text, and offers low, medium, high and xhigh reasoning-effort settings.

For prompts below 200,000 tokens, SpaceXAI lists standard rates of $2 per million input tokens, $0.50 per million cached input tokens and $6 per million output tokens. Above that threshold, the rates rise to $4, $1 and $12, respectively. The Grok 4.7 Fast option uses the same model at twice the standard token rates and is available in Cursor and Grok Build, but not through the public SpaceXAI API.

Cursor lists Grok 4.7 at the same standard on-demand rates and the Fast option at $4 per million input tokens, $1 per million cached input tokens and $12 per million output tokens. Cursor applies its long-context surcharge above 256,000 input tokens, compared with SpaceXAI’s 200,000-token threshold.

SpaceXAI says Grok 4.7 uses a larger base model than Grok 4.6, underwent a longer reinforcement-learning run weighted toward multi-hour work, and improved self-verification and long-context management. Those development and qualitative-performance claims were not independently validated in the opened source set.

A CursorBench 4.0 result provides one workload comparison: Grok 4.7 Extra High scored 46.3% at an average cost of $6.01 per task, compared with 41.4% and $6.10 for Grok 4.6 Extra High. Several Fable 5.1 configurations and Opus 5 Max scored higher in the same table, so the result does not establish broad model superiority. SpaceXAI’s wider price-performance comparisons likewise depend on vendor-selected benchmarks and methods.

SpaceXAI also calls Grok 4.7 its strongest model for refusals and jailbreak resistance. The company says its HackerBench v0.3 test allowed 3.3% of risky dual-use prompts through, but it describes HackerBench as its own benchmark. Those results do not independently establish the broader safety claim.