SpaceXAI has released Grok 4.7, a new AI model for programming and complex workloads that can take hours. In some tests, it already outperforms Claude Fable 5.1, and the base API fee starts at $2 per million input tokens. However, for very long contexts, the fee doubles.
SpaceXAI has released Grok 4.7, a new AI model for programming and complex workloads that can take hours. In some tests, it already outperforms Claude Fable 5.1, and the base API fee starts at $2 per million input tokens. However, for very long contexts, the fee doubles.
SpaceXAI reports on the launch.
The company calls Grok 4.7 its strongest model for programming and intellectual work — such as information analysis, document work, and research.
To “pump” the new version, SpaceXAI took a larger base model and trained it for longer on complex tasks, where it is not enough to simply give one answer. The model had to go through several steps on its own: understand the task, choose the necessary actions, perform them and check the result. So the developers claim that Grok 4.7 now copes better with such long chains of work and notices its own mistakes more often.
This is also noticeable in tests where the AI has to do similar things: figure out the task on its own, work with files, and run the necessary commands. For example, in Terminal-Bench, Grok's result almost doubled — from 20.3% to 38%. And in CursorBench 4.0, where the models perform long programming tasks, — from 40.4% to 46.3%.
However, Claude Fable 5.1 is still stronger in such tasks. In the same Terminal-Bench, it scored 57.9%. But in individual tests on programming, electrical engineering and legal tasks, Grok shows results on par with or higher than Claude. It is also worth noting that so far these are benchmarks published by SpaceXAI.
Grok 4.7’s base API pricing remains the same as Grok 4.6: $2 per million inbound and $6 per million outbound tokens. Claude Fable 5.1 costs $10 and $50, respectively. However, Grok 4.7 has a 500,000 token cap, and if the request exceeds 200,000 tokens, SpaceXAI doubles the fee to $4 per million inbound and $12 per million outbound tokens.
You can try Grok 4.7 now via the SpaceXAI API, Grok Build, OpenRouter, Vercel, and Cloudflare, as well as in Cursor. There is also a separate Fast version: it responds about twice as fast, but also costs twice as much. For now, it is only available in Cursor and Grok Build.
Previously, dev.ua wrote that SpaceX plans to train Grok using the company's internal data, including the work of its employees. According to Musk, this should help the AI better perform real-world work tasks.
