Grok 4.7: terminal and web-agent coding gains, lower scores across Vals AI's suite
The Aggregate Digest — Tuesday, September 22, 2026
Grok 4.7: terminal and web-agent coding gains, lower scores across Vals AI's suite.
Grok 4.7 enters the table at #16, three rating points above Grok 4.6 and inside its error bar. Whether it is an upgrade depends on who ran the test. The official Terminal-Bench and Next.js agent boards show large gains. Vals AI and Mercor score it below Grok 4.6 on professional tasks. Best available model stays GPT-6 at 2118.
The Aggregate fitted ranking, September 22 snapshot. Grok 4.7 carries a 16-point standard error on 132 results; the dot axis is zoomed.
Terminal agent work nearly doubles. Terminal-Bench 4.0 gives an agent 66 tasks to finish inside a terminal, most of them scientific or engineering work. On its official board Grok 4.7 scores 37.6 percent to Grok 4.6's 20.3 and is sixth of 15. GPT-6 leads at 58.2 and Claude Fable 5.1 has 57.9.
tbench.ai Terminal-Bench 4.0 official leaderboard, September 22 snapshot. Eight of 15 entries shown; percent of tasks passed.
First on Vercel's Next.js agent board. Next.js Agent Evals run coding agents on 31 real Next.js tasks, scored pass@4. Run through OpenCode without bundled docs, Grok 4.7 passes 94 percent, first of 12, ahead of Gemini 3.8 Flash at 90. Grok 4.6 scored 71. With Next.js docs bundled in AGENTS.md it stays at 94 while three models reach 97.
Vals AI and Mercor measure a step backwards. The Vals Index combines Vals AI's finance and coding benchmarks, weighted by each sector's share of US GDP. Grok 4.7 scores 54.15 percent to Grok 4.6's 59.17, 24th of 56 where Grok 4.6 is 14th. Excel Modeling falls from 62.73 to 54.94 and Code Migration from 44.57 to 34.7. Vals' own Terminal-Bench 4.0 run goes the opposite way from the official board, 12.12 to Grok 4.6's 17.17. Mercor APEX, long professional workflows with tools, drops from 65.3 to 54.6. Vals' stored rows do not name a reasoning setting.
September 22 results from tbench.ai, Vercel's Next.js Agent Evals, Vals AI and Mercor. Each row compares the same task and metric across the two models. Different tasks, not a combined score; five rows come from one evaluator.
Workplace automation edges up in Artificial Analysis' testing. At extra-high effort, AutomationBench-AA Strict, which requires full success on workplace objectives under guardrails, scores 37.75 percent to 34.55 for Grok 4.6 at the same setting, third of 171 behind two GPT-6 configurations. The guardrail-adjusted headline score is lower, 65.56 to 67.0.
Terminal and web-agent coding is where the evidence says to try it. The Vals losses stay unsettled until the reasoning setting behind them is known.
