Grok 4.7: Inexpensive API meets high token consumption

Grok 4.7 is now available for developers and, according to SpaceXAI, is significantly more powerful, especially in coding, agentic tasks, and knowledge work. It follows Grok 4.6 by about a month and is based on a new, larger base model. In subsequent reinforcement learning, SpaceXAI trained the model longer and with a more demanding task mix. This is intended to help Grok 4.7 work on difficult tasks for longer, better review its output, and manage long contexts more effectively. For improved coding and agent capabilities, SpaceXAI also used anonymized workflow data from Cursor, which has belonged to SpaceX since August.

According to its statements, SpaceXAI has also revised the protection mechanisms. The company reports better results in tests for jailbreak resistance as well as cyber and biosafety.

The model is available via the xAI API, SpaceXAI’s coding agent Grok Build, and the Cursor development environment, as well as via OpenRouter, Vercel, and Cloudflare. The regular Grok offerings on grok.com, in the mobile apps, and on X are expected to follow later, according to the Model Card.

Significant Leap in Agentic Work

In the Artificial Analysis Intelligence Index, Grok 4.7 achieves 46 points with the highest reasoning level, xHigh, only two more than Grok 4.6. According to Artificial Analysis, this places SpaceXAI among the top four AI labs, but with Grok 4.7, it remains behind the current top models Claude Fable 5.1 and GPT-6 Astra in the index. The comparatively small jump from Grok 4.6 is reflected in the individual tests: outside agentic knowledge work, the progress remains manageable overall, according to independent testers.

However, the gain is significantly larger there: in the AA Briefcase benchmark, which tests realistic, complex work tasks over extended periods, Grok 4.7 is only slightly behind Claude Opus 5 and Claude Fable 5.1. On GDPval-AA, which evaluates the quality of practical work results from various professional fields, Grok 4.7 also reaches top-tier levels.

The model also improves in coding. Grok 4.7, in conjunction with Grok Build, achieves 56 points in the Coding Agent Index, nine points more than Grok 4.6. Among models in their respective native agent environments, the combination ranks fourth, behind Claude Fable 5.1, GPT-6 Astra, and Claude Opus 5. In DeepSWE, the success rate increases from 65 to 73 percent; in Terminal-Bench 4.0, it rises from 18 to 33 percent.

Inexpensive Tokens, High Consumption

Regarding pricing and basic technical specifications, nothing has changed compared to Grok 4.6: Grok 4.7 costs two dollars per million input tokens and six dollars per million output tokens for prompts under 200,000 tokens. Cached input tokens are charged at 50 cents per million. Above 200,000 prompt tokens, all three rates double. The context window remains at 500,000 tokens, with four reasoning levels available from Low to xHigh. A faster, more expensive Fast variant is currently exclusively available via Cursor and Grok Build.

Artificial Analysis measures an average of around 81,000 output tokens per task in the Intelligence Index for Grok 4.7 with the highest reasoning level, xHigh. Grok 4.6 xHigh requires around 38,000 there, and GPT-6 Astra Max approximately 27,000 tokens.

The high token consumption is also reflected in the costs per task: the testers estimate them at xHigh, considering input, cache, reasoning, and output tokens, to be about 3.74 dollars. GPT-6 Astra Max, despite significantly higher token prices, comes in at around 3.26 dollars. Thus, the low price per token does not automatically lead to low costs per completed task. Especially in productive agent workflows, token consumption is crucial alongside model performance.

(dmk)