Grok 4.7 buys 5.9 points of CursorBench at Grok 4.6's token price

The score moved 5.9 points and the rate card did not move at all, so on the standard tier xAI now charges about 12.7% less per point of CursorBench 4.0 than it did for Grok 4.6 ^[1][2][12]. That is the seller's version. The buyer's version is cost per finished task, and it runs the other way if the model does what xAI describes, which is stay with demanding assignments for longer and review its own output more carefully ^[3]. Persistence is output tokens. At $6 per million, a run that emits 15% more tokens bills 15% more ^[12][8].

The capability came out of the training budget. Grok 4.7 sits on a larger base model than 4.6 and got a longer reinforcement-learning run, weighted toward problems that can take many hours to finish instead of short prompts ^[4][5]. xAI did not publish what that run cost. The company frames the release as a practical upgrade over 4.6 and not a change in pricing or serving speed ^[2].

The gains are uneven. Terminal-Bench 4.0 went from 20.3% to 38.0% on xAI's reported harness, an 87% relative move ^[8][3]; EEBench went from 53.0 to 64.0, or 20.8% ^[9][4]; CursorBench 14.6% ^[1]; DeepSWE v1.1 from 65.2% to 71.0% at high effort, 8.9% ^[7][7]; and AA Briefcase v1.1, the multi-hour office test, from 1,546 to 1,657, or 7.2% ^[10][5]. The account of the launch says those results put Grok 4.7 near other frontier systems on some tests and behind them on others, depending on the evaluation and the effort setting ^[11].

Context stays at 500,000 tokens, with text and image input, tool use, search and code execution ^[15]. Filling that window once costs a dollar at $2 per million input tokens ^[6]. The model was also trained to work natively with the Grok Bot harness, which xAI expects to help with conversation, document drafting and general office-style work ^[17].

Three ways this goes. If output per completed task holds flat, the unchanged rate card is a real price cut and the 12.7% per point is what buyers collect ^[2]. If the extra persistence pushes output per task up by more than the 14.6% score gain, buyers are paying more for better work ^[1]. And since 4.7 landed only weeks after 4.6 ^[18], the rate card has a decent chance of moving before the next benchmark table does. I'd expect the middle case, since self-checking and long-horizon persistence are both token-spending behaviours ^[3]. The check is one repository task run on both models, with the comparison made on output token counts instead of scores.