Grok 4.7 is cheap on paper, but the price doubles for long tasks

On September 21, 2026, xAI released its new AI model, Grok 4.7. According to xAI, it is a major upgrade for coding and knowledge work. The company says the model was trained with a longer reinforcement learning run (training in which the model is rewarded for good answers) on a harder mix of tasks and has become better at checking its own work. It is available through Cursor, Grok Build, the Grok API and third-party platforms.

Price depends on task length

The list price in the API (the access developers use to build the model into their own programs) is $2 per million input tokens and $6 per million output tokens. A token is a small piece of text, typically a word or part of a word. Cached input, meaning text the model has seen before, costs $0.50 per million tokens.

But once the prompt reaches 200,000 tokens, the entire request is billed at a higher rate: $4 per million input tokens, $1 for cached input and $12 per million output tokens. That means the price doubles within the model's 500,000-token context window. According to Mervin Praison, xAI's announcement mentions only the low list price, while the more expensive tier appears in the API documentation. A faster variant costs twice as much and delivers output roughly twice as fast.

xAI's own numbers and the independent ones

xAI itself reports these results for Grok 4.7: 46.3 percent on CursorBench 4.0, 71.0 percent on DeepSWE v1.1, 64.0 percent on EEBench, 19.6 percent on Harvey Legal Agent and 56.7 percent on HealthBench Professional. According to Mervin Praison, Grok trails Fable 5.1 on CursorBench and GPT-5.6 Sol on HealthBench and DeepSWE, while it leads on EEBench and Harvey Legal Agent among the models shown. These are the vendor's own figures.

Vals AI has compared Grok 4.7 with Claude Opus 5.5. According to Vals AI, Claude scores higher on 20 of 22 shared benchmarks, while Grok leads on two. The averages were:

- Coding: Claude Opus 5.5 67.13 vs. 43.60 percent for Grok 4.7 - Science: Claude 58.65 vs. 34.83 percent for Grok - Finance: Claude 60.03 vs. 51.15 percent for Grok - Health: Claude 70.61 vs. 69.47 percent for Grok

The largest gap was 74 points on ProofBench v1.1, in Claude's favor. Vals AI writes that the confidence intervals overlap on only 4 of 21 comparable benchmarks, so the differences are mostly statistically meaningful.

On the other hand, Grok was faster and cheaper in the Vals Index test: 35 minutes and $12.12 versus 1 hour and 19 minutes and $32.14 for Claude Opus 5.5. xAI itself places Grok 4.7 at the frontier for price and performance on long coding tasks.

Caveats

According to Let's Data Science, the published benchmarks compare Grok 4.7 at the highest reasoning level (xhigh) with Grok 4.6 at a lower one (high). Teams considering the model should therefore run their own tests with identical settings. It is also not clear from the Vals AI page which settings Grok was run with.

What it means for you

The news is most relevant if you build AI into your own programs or use coding tools like Cursor, where Grok 4.7 can now be selected. If your tasks are short, the list price is accurate, but if you feed in very long texts or codebases, the price doubles once you pass 200,000 tokens. So test the model on your own tasks with the same settings as the competitors before choosing it. If you need the best possible result, Claude Opus 5.5 scores higher in the independent test, while Grok was faster and cheaper in the same test.