Source briefing. Provider claims are attributed to the sources below. This page does not claim a new independent DARC benchmark.
For: Developers and teams putting AI into practical use
⚡ TL;DR — the 30-second version
- Use grok-4.7, not the old Grok 4 entry.
- Standard API rates: $2 input / $6 output per million tokens.
- Reported DeepSWE uses high effort; the other listed coding results use xhigh.
- No open-weight claim is made for Grok 4.7.
Grok 4.7 was introduced on 21 September. xAI describes a larger base model, a longer reinforcement-learning run and harder long-running tasks than Grok 4.6. It is available in the Grok API, Grok Build and Cursor; the current API catalog lists a 500,000-token context window.
Standard input/output prices are $2 and $6 per million tokens. xAI also offers a fast variant at twice the price. The launch page's speed and cost claims are provider comparisons, not measurements made on this site.
“A stronger model with unchanged pricing is worth testing. It is not permission to skip your regression set.”
The release reports 46.3% on CursorBench 4.0 and 37.6% on Terminal-Bench 4.0 at xhigh effort. Its 71.0% DeepSWE v1.1 result is explicitly marked high effort. Preserve that exception when copying the table. A percentage without benchmark version and run settings is not a useful comparison.
For a migration, test long tasks, verification behavior, context handling and the shape of tool calls. Also measure human intervention and rejected changes. Provider benchmark improvement can be meaningful while still failing to predict the cases unique to your codebase.
Field guide / 01
Keep the footnotes attached to the score
The current release reports several coding and professional-work evaluations. Do not copy a high-effort DeepSWE number into a table labeled xhigh, or treat a professional-work Elo as a percentage. The measurement's unit and run conditions are part of the result.
Use the launch table to identify tasks worth testing locally. A codebase-heavy workload should include dependency changes, failing tests and verification steps, not only isolated function writing.
Field guide / 02
A practical adoption test
Freeze two or three representative issues and give the candidate the same tools and resource limits as your baseline. Require a complete action trace and a final list of unverified assumptions. Review patches without being told which model produced them where feasible.
Record provider errors, timeouts and recovery. The ability to work longer is useful only if you can stop the run, preserve the result and control what tools it can reach.
Before you ship
0/5Use this as a review aid. Checks reset when you leave this page.
Questions that come up
Is the launch table a DARC evaluation?
No. It is a provider publication. DARC links and attributes those values and does not imply an independent rerun.
Sources & verification
Last checked 23 Sept 2026. Publication dates and access conditions belong to the linked originals.
Editorial policy and correction log →