Grok 4.7 Benchmarks: Vendor Tables, Independent Runs, and Harness Effects
Grok 4.7 has different benchmark results depending on evaluator, harness and effort setting, so provenance must be preserved instead of collapsing scores into one leaderboard.
Key takeaway
SpaceXAI reports 46.3% CursorBench 4.0, 71.0% DeepSWE v1.1 at high effort, 64.0% EEBench in the launch table, 1,657 AA Briefcase Elo, 37.6% Terminal-Bench 4.0, 19.6% Harvey Legal Agent Benchmark and 56.7% HealthBench Professional. Artificial Analysis independently reports 46 Intelligence Index, 1,657 AA-Briefcase, 1,695 GDPval-AA and 56 Coding Agent Index with Grok Build.
Updated on Sep 26, 2026
01
Abstract
Grok 4.7 benchmark results illustrate why frontier-model evaluation requires provenance. Vendor tables, model-card evaluations, and independent measurements can differ because model settings, harnesses, versions, and evaluators differ.
02
Introduction
A benchmark result measures more than model weights. The surrounding harness, tools, retry policy, effort level, evaluator, and benchmark version can materially change the observed result.
03
Problem
Grok 4.7 has multiple published measurements for some benchmark families. Treating all values as one leaderboard score can therefore produce misleading comparisons.
04
Background
A useful benchmark record can be represented as model, benchmark version, effort, harness, evaluator, score, and date. Scores should only be directly compared when relevant dimensions align.
05
Methodology
The analysis keeps official launch results, model-card evaluations, and independent Artificial Analysis measurements as separate evidence ledgers. Vendor and independent numbers are not averaged.
Grok 4.7 benchmark / cost visual
Distinct benchmark-cost visual that includes a Grok 4.7 xhigh point. Third-party social-media image; not an official xAI benchmark graphic.
Wes Roth on X; third-party asset
06
Architecture
The measured system includes the model plus its evaluation harness, tools, permissions, retries, effort configuration, and grading system.
07
Dataset
The benchmark set includes CursorBench, DeepSWE, EEBench, Artificial Analysis Briefcase, Terminal-Bench, GDPval, HealthBench Professional, and other named evaluations.
08
Training
No new training procedure is introduced by this benchmark analysis. Important variables occur at inference time, including reasoning effort, harness configuration, retry policy, tool availability, and evaluator implementation.
09
Experiments
10
Baselines
The published comparison set includes Grok 4.6, GPT-5.6 Sol, Fable 5.1, and independent Artificial Analysis measurements. Comparisons are retained with their original effort and harness conditions.
11
Results
Artificial Analysis reports an Intelligence Index of 46 and Coding Agent Index of 56 using Grok Build, with AA-Briefcase 1657, GDPval-AA 1695, Terminal-Bench 33%, DeepSWE 73%, and SWE-Atlas-QnA 63% in its independent measurements.
12
Ablation
The key measurement factors are effort level, harness, evaluator, benchmark version, retry policy, and tool configuration. Changing these factors can produce different scores even when the underlying model remains unchanged.
13
Limitations
Vendor and independent protocols are not always identical, benchmark pages can change, live evaluations may use different harness versions, and some results use different effort settings.
14
Conclusion
Grok 4.7 should be evaluated through a provenance-aware benchmark matrix rather than one aggregate number. Holding harness, effort, tools, retries, and evaluator constant enables stronger like-for-like comparisons.
15
References
Grok 4.7 official launch material and model card; Artificial Analysis independent evaluations; benchmark-specific documentation for CursorBench, DeepSWE, EEBench, Terminal-Bench, GDPval, HealthBench, and related evaluations.
Source trail
Primary papers, implementation pages, and external source records used to ground this explainer.
Introducing Grok 4.7
Official launch announcement with training changes, benchmark comparison, pricing, availability and safety claims.
Grok 4.7 model card
Official benchmark and safety evaluation details, including additional model-card runs.
Benchmarking Grok 4.7
Independent measurements including Intelligence Index, Coding Agent Index, token use and task duration.
Official Grok 4.7 visual set
The official launch page contains the published launch artwork plus benchmark/professional-work and cost-effectiveness visuals. The verified page fetch did not expose stable direct image asset URLs, so no unverified figure URL was inserted.
01SpaceXAI: Introducing Grok 4.702Grok 4.7 model documentation03Grok 4.7 model card04Artificial Analysis: Benchmarking Grok 4.7