Grok 4.7’s coding score is 38% on xAI’s table and 26% at Artificial Analysis
Vlad Makarov, Traictory, 22 September: twelve points on Terminal-Bench 4.0, one test, two harnesses. The $2/$6 price, unchanged from 4.6, is the part that holds. AA’s Intelligence Index has Grok 4.7 at 46, mid-pack against Fable 5.1 and GPT-6 Astra at 53.
Feature2 minIn-depth
Maya Chen2 min read
- Grok 4.7
- Artificial Analysis
- Terminal-Bench
- xAI
- Benchmarks

Photo: TraictorySource page
The morning launch post is already on this site. What Tuesday added is a second reading of the same coding benchmark. Vlad Makarov at Traictory put xAI’s Terminal-Bench 4.0 row — 38.0 percent at the xHigh setting — next to Artificial Analysis’s 26 percent on the same 66-task set. Twelve points. xAI did not name the harness behind its number. Artificial Analysis runs all 66 tasks with mini-swe-agent and reports pass@1 averaged over three repeats. Terminal-Bench 4.0 is a Laude Institute / Stanford / open-source suite; the 4.0 cut dropped eight saturated or refusal-prone tasks. A different scaffold and a different effort budget routinely move agentic scores by double digits. Makarov’s line is the right one: neither figure is a lie, and neither is published in a form a reader can reconcile.
The price half of Monday’s pitch holds. Grok 4.7 is $2 per million input tokens and $6 per million output, the same as 4.6, against $4/$20 for GPT-5.6 Sol and $10/$50 for Claude Fable 5.1. A fast variant doubles output speed and doubles that price. Artificial Analysis, as Traictory quotes it, measured the default xhigh configuration at 39.3 output tokens per second — low for its tier, where the median is 72.5 — and called the model “notably slow and very verbose,” 240 million output tokens on the Intelligence Index against a 94 million median. Time to first token, 0.88 seconds, is competitive. The cheap configuration the scores were run on is not the fast SKU. “Twice as fast at half the price” describes two products.
Mid-pack on the independent index
On Artificial Analysis Intelligence Index v4.3.2, ten benchmarks, Grok 4.7 scores 46. Fable 5.1 and GPT-6 Astra lead at 53. AA notes that Grok’s two highest reasoning levels “appear to perform about the same,” which undercuts the top of the effort range. On Terminal-Bench, Traictory adds, the cheaper DeepSeek V4.1 Flash scores 27 percent — a point above AA’s Grok reading. xAI’s own launch table already had Grok losing Terminal-Bench 4.0 to Fable 5.1 (57.9 percent) and CursorBench 4.0 to Fable (51.8 against 46.3). Those rival columns are xAI-measured too. The morning story treated $4.69 a CursorBench task, LatchBio 62.4 percent and HackerBench 3.3 percent as vendor yards. This one does not rerun them.
What would close the gap is ordinary: name the harness and the repeat policy, or rerun Terminal-Bench 4.0 on the public mini-swe-agent so the two 38 and 26 readings become one table. Until then the buy is the tariff. Grok 4.7 is still $2/$6. It is not, on the independent board, a capability lead.


