Grok 4.7 Tops VulcanBench Frontier v4 at 93.15 While Scoring 37.6% on Terminal-Bench 4.0
Grok 4.7 from xAI scored 93.15 on VulcanBench Frontier v4 at the xhigh effort level, claiming the top position on that leaderboard in independent testing conducted between October 1 and 3, 2026. On Terminal-Bench 4.0, the same model scored 37.6%, placing it below Claude Sonnet 5.5 (70.6%), GPT-6 Sol (43%), and Claude Haiku 5.5 (39.2%).
Two Benchmarks, Two Stories
VulcanBench Frontier v4 and Terminal-Bench 4.0 test different aspects of model capability. The 55-point gap between Grok 4.7’s results on the two evaluations illustrates a pattern that has emerged across frontier model testing in 2026: no single model dominates every benchmark category, and rankings depend heavily on the task distribution each evaluation emphasizes.
Grok 4.7 was released September 21, 2026. xAI described it as using a larger base model than Grok 4.6 with a longer reinforcement learning training run, at identical pricing: $2 per million input tokens and $6 per million output tokens, with a 500,000-token context window.
Benchmark Comparisons
Grok 4.7 results across five agentic evaluations:
| Benchmark | Score | Notes |
|---|---|---|
| VulcanBench Frontier v4 | 93.15% | Independent, Oct 1-3, #1 on board |
| DeepSWE v1.1 | 71.0% | High effort, vendor-reported |
| EEBench | 64.0% | Vendor-reported |
| CursorBench 4.0 | 46.3% | Vendor-reported |
| Terminal-Bench 4.0 | 37.6% | Vendor-reported |
The VulcanBench and DeepSWE results suggest strong performance on code transformation and autonomous repair tasks. The Terminal-Bench 4.0 result, which measures a different mix including shell navigation, system interaction, and multi-step terminal workflows, shows a substantially lower relative position.
For context, Terminal-Bench 4.0 comparisons across the frontier:
- Claude Sonnet 5.5: 70.6%
- GPT-6 Sol: 43.0%
- Claude Haiku 5.5: 39.2%
- Grok 4.7: 37.6%
- GPT-6 Luna: 13.0%
Terminal-Bench 4.0 Is a Hard Reset
Terminal-Bench 4.0 is a harder evaluation than its predecessors. When Terminal-Bench 3.0 launched in August 2026, frontier models that had scored in the mid-80s on Terminal-Bench 2.0 dropped to the mid-30s. The current version continues that pattern: models that excel on code generation tasks tend to struggle on the workflow-heavy terminal evaluation.
Grok 4.7’s predecessor, Grok 4.6, scored 95.6% on SWE-bench Verified in August 2026, placing it fourth on that leaderboard. Grok 4.7 does not yet have an independent SWE-bench Verified result. xAI’s vendor-reported benchmark coverage for Grok 4.7 leans toward DeepSWE and CursorBench, which the company highlighted in launch materials, with Terminal-Bench 4.0 included but unemphasized.
Enterprise Implications
For buyers comparing Grok 4.7 against alternatives for terminal automation and shell-based agentic workflows, Terminal-Bench 4.0 is the more relevant signal. A 30-point gap — Grok 4.7 at 37.6% versus Sonnet 5.5 at 70.6% — is large enough to affect task completion rates in production.
On VulcanBench-aligned tasks, the picture reverses: Grok 4.7 leads the leaderboard in independent evaluation. The divergence makes Grok 4.7 a benchmark-specific choice for agentic deployments rather than a universal capability upgrade over Grok 4.6.
AA Briefcase v1.1 rates Grok 4.7 at an ELO of 1657, a useful cross-benchmark sanity check that places it within the frontier tier but below the Opus 5.5 and GPT-6 Astra positions on the same index.