GPT-6 leads Claude Fable 5.1 by eight points on the table, and on the long-running agent tasks Vals AI published this week the gap is far wider
The Aggregate Digest — Sunday, September 20, 2026
GPT-6 leads Claude Fable 5.1 by eight points on the table, and on the long-running agent tasks Vals AI published this week the gap is far wider. If the job is an agent working for hours, that is the number to read. Vals AI's Time Horizon Index has 11 entries, and GPT-6 at max effort scores 90.5 percent progress there to 63.33 for the max-effort run of Claude Fable 5.1; the third entry, GPT-5.6 Sol at max effort, is at 23.83. On Terminal-Bench Science the same pair goes 65.71 percent of tasks resolved to 34.29, and on Terminal-Bench 4.0, 66 containerised terminal tasks weighted toward science and frontier engineering work, GPT-6 is first of 28 at 57.07 to 49.49. FrontierSWE V2 gives an agent 34 coding tasks with a 20-hour budget each; GPT-6 scores 65.5 to 56.3 for Claude Fable 5.1, with GLM-5.3 third at 30.2. Scale's Remote Labor Index went the same way, 20.83 to 17.92. GPT-6 holds best available model at 2113 over Claude Fable 5.1 at 2105, and Claude Mythos 5 holds best measured at 2135.
The shorter and more professional the task, the less of that gap remains. Claude Fable 5.1 leads the Vals Index, which weights finance and coding tasks together, at 68.83 percent. Claude Opus 5 is second at 67.21 and GPT-6 third at 66.61. Claude Fable 5.1 also leads the 58 entries on Excel Modeling at 76.67, where GPT-6 is sixth. Legal Research Bench is research tasks across US law, and there Claude Fable 5.1 shares the lead at 55.29 percent with Claude Opus 5 and with Muse Spark 1.3 at max effort, while GPT-6 is 20th of 61. GPT-6 is 22nd of 61 on Finance Agent v2, a board Gemini 3.8 Flash leads. Vibe Code Bench has agents build an application from a product prompt, and Claude Fable 5 leads its 96 entries at 90.35 percent with GPT-6 third at 89.59. The new Vibe Code Bench 1-100 takes a passing app and extends it through up to ten dependent requests, and every model drops below 30 percent: Claude Opus 5 28.53, then Claude Fable 5.1 at 28.0 and GPT-6 at 27.64.
Games split the same way. GPT-6 at max effort is first or tied first on five of BALROG's six boards, where models play games such as Crafter and NetHack as autonomous agents, and seventh of 37 on TextWorld. Kaggle's Game Arena belongs to other models this week. Gemini 3.8 Flash took the chess board from Gemini 3.1 Pro (Preview), with Gemini 3.7 Flash second and GPT-6 eighth of 31, and Claude Opus 5 took Four in a Row with Claude Fable 5.1 second and GPT-6 tenth of 32.
Elsewhere. Cooper Labs' Insurance Agent Benchmark asks one question each of 166 real commercial insurance documents, graded against a human-verified key, and publishes every model twice. With Cooper's own software reading the documents first, Gemini 3.8 Flash leads 17 entries at 85.6 percent and Claude Opus 5 is at 85.3. Sent the file and the question alone, Grok 4.6 leads at 77.6, Gemini 3.8 Flash is sixth at 75.9 and Claude Opus 5 ninth at 73.4. Kimi K3 remains the best open-weights model at 1974.
