SpaceXAI released Grok 4.7 on September 21, pitching the model for coding and longer knowledge-work tasks while keeping its standard API token prices level with Grok 4.6. The company says it is available in Cursor and Grok Build, through the Grok API, and on several model gateways. Its developer documentation lists the API model as grok-4.7.
The company says Grok 4.7 uses a larger base model and a longer reinforcement-learning run focused on harder, extended tasks. It also says the model was trained to work with the Grok Bot harness. These are descriptions of SpaceXAI's training and performance claims, not findings from hands-on testing by Digg.
The benchmark gains need their settings
In SpaceXAI's launch table, Grok 4.7 scores 46.3% on CursorBench 4.0 versus 40.4% for Grok 4.6, and 38.0% versus 20.3% on Terminal-Bench 4.0. The same table lists an AA Briefcase v1.1 score of 1,657 for 4.7 against 1,546 for 4.6. The comparison has a significant caveat: the table labels Grok 4.7 as running at xhigh reasoning effort and Grok 4.6 at high. The scores are company-published results under those stated settings, so they should not be read as a controlled measure of the model upgrade alone.
SpaceXAI also claims stronger refusal and jailbreak resistance from a new safeguard system. Those safety figures come from its announcement and have not been independently checked here.
Price and availability
The standard model costs $2 per million input tokens and $6 per million output tokens, matching the prices SpaceXAI lists for Grok 4.6. The documentation specifies a 500,000-token context window and text and image input with text output. SpaceXAI says a faster version costs twice as much; its documentation says that variant is limited to Cursor and Grok Build and is unavailable through the public API. The regular model is available through the API now.
Comments