TUE, SEPTEMBER 22, 2026
Independent · In‑Depth · Practitioner‑Tested
✎ Large Language Models

Grok 4.7 Is Out at $2 In and $6 Out. The Benchmarks Are Uneven.

SpaceXAI shipped Grok 4.7 on 21 September at $2 per million input tokens and $6 per million output, with a fast variant at double both. It scores 71.0% on DeepSWE v1.1 and 46.3% on CursorBench 4.0, but only 38.0% on Terminal-Bench 4.0 and 19.6% on the Harvey legal agent benchmark. Here is what those numbers mean for your workload.

By AIToolsRecap September 22, 2026 6 min read 35 views
Home Articles Large Language Models Grok Grok 4.7 Pricing and Benchmarks: Where It Wins ...

SpaceXAI released Grok 4.7 on 21 September 2026. The pitch is explicit: twice as fast at half the price of comparable models. The pricing supports that. The benchmarks support it in some places and not others.

What It Costs

Standard Grok 4.7 runs at $2 per million input tokens and $6 per million output tokens. There is a fast variant at $4 in and $12 out that trades double the price for roughly double the output speed.

For comparison against the models it is aimed at, that output price is the part that matters. Most agentic coding work is output-heavy, and $6 per million output is materially below frontier pricing from OpenAI and Anthropic.

The caveat several reviewers raised on launch day: Grok 4.7 consumes a lot of tokens per task. A lower per-token price does not automatically mean a lower per-task bill. If you are evaluating it, measure cost per completed task rather than cost per million tokens.

The Benchmark Numbers

SpaceXAI published these on the launch page:

  • CursorBench 4.0: 46.3%
  • DeepSWE v1.1: 71.0% at high effort
  • EEBench: 64.0%
  • Terminal-Bench 4.0: 38.0%
  • HealthBench Professional: 56.7%
  • Harvey Legal Agent Benchmark: 19.6%
  • LatchBio biosafety: 62.4%
  • AA Briefcase v1.1: 1,657
  • GDPval: 1,695 Elo
  • HackerBench v0.3: 3.3% risky refusal rate

Reading the Spread

DeepSWE at 71.0% is a strong software-engineering result and the number SpaceXAI leads with. Terminal-Bench at 38.0% is a much weaker showing on long-horizon shell work, and the two together describe a model that is good at bounded code tasks and less reliable when it has to drive a terminal for an extended session.

The 19.6% on the Harvey legal benchmark is the outlier. That benchmark measures multi-step legal agent work, and a sub-20% score means this is not a model to point at contract review or discovery without heavy supervision. HealthBench Professional at 56.7% sits in the middle.

The pattern across all ten: Grok 4.7 is priced and tuned for coding and general knowledge work. The domain-specific professional benchmarks are where the gap to Claude and GPT-class models shows up clearly, and SpaceXAI did not hide them.

Where You Can Use It

Available at launch through Cursor, Grok Build, the Grok API directly, third-party coding harnesses, and the usual model routers and cloud platforms. If you already run Cursor, it is a dropdown change.

Who Should Switch

If your workload is high-volume code generation and your current bill is driven by output tokens, run a week of real tasks against it and compare total cost per merged PR, not per token. The 71.0% DeepSWE result says the quality is there for that use case.

If you are doing legal, clinical or compliance work, the published numbers say stay where you are. A 19.6% agent benchmark is not a rounding error.

If you are on the fence, the fast variant at $4 and $12 is still under a lot of frontier pricing, and doubling speed on an agentic loop changes the feel of the tool more than most people expect.

Sources

Tags
AI NewsGrokCoding AI2026

Spot an inaccuracy?

We verify facts before publishing and correct errors promptly. If something in this article is wrong or outdated, let us know.

Report an error →