DeepSeek V4.1 Flash is the best open-weights model after Kimi K3 two days after arriving, and it beats GPT-6 on AutomationBench-AA
The Aggregate Digest — Sunday, September 13, 2026
DeepSeek V4.1 Flash is the best open-weights model after Kimi K3 two days after arriving, and it beats GPT-6 on AutomationBench-AA. That is a 156-entry board where agents pursue workplace objectives inside guardrails, and DeepSeek V4.1 Flash at max effort completes 68.89 percent of the objectives to 68.49 for GPT-6 at max effort. The same model tops both LiveBench coding boards of 57 entries: Python at 90 with the next score at 75 from Claude Opus 5 running at max effort, and JavaScript at 81.82. Vals AI SkillsBench asks models to work from skill files on office and document tasks, and DeepSeek V4.1 Flash leads its 34 entries at 69.8 percent to Grok 4.5's 66.03. Its rating of 1955 rests on 150 results. That puts it 18th among available models and 23 points behind Kimi K3 at 1978.
The strict boards are where it slips. When AutomationBench-AA counts only fully completed tasks it drops to fourth at 35.46 percent behind three GPT-6 runs. Its lab-submitted 63.9 on Humanity's Last Exam, tools allowed, is the top of 26 such scores and none of them has been checked independently. Benchmarks.bio's SpatialBench has models build analysis workflows on spatial transcriptomics data, and there DeepSeek V4.1 Flash is second of 23 at 75.36 percent behind Grok 4.6 at 77.39.
Gemini 3 Deep Think enters the top 20 of available models at 20th with a rating of 1943 on 37 results, the thinnest record in the top 20. Hard mathematics and competition code are what it does well. It is first of 58 on LiveCodeBench Pro, competitive programming problems that 99.9 percent of competitive coders cannot solve, ahead of Gemini 3.1 Pro (Preview), and third of 12 on IMO-Bench's proof writing at 76.7 percent behind Aletheia at 91.9 and GPT-5.5 Pro at 88.1. IUMB is 16 introductory undergraduate math problems, and there it is sixth of 55 at 87.5 percent where two entries score 100. Ten of its 37 results come from Google's own Gemini 3 Deep Think evaluation page, whose boards have two to four entries each with Gemini 3 Deep Think on top. Another 16 are M3-Bench sub-boards where it sits mid-field. TrackingAI's IQ Test gives it 64.71 percent where Claude Fable 5 scores 98.04.
GPT-6 is the best available model at 2110, six points ahead of Claude Fable 5.1 at 2104, and it keeps adding boards. Vending-Bench 2 simulates a year of running a vending machine business, and GPT-6 leads its 63 entries with a closing balance of $15,514.70 to Claude Opus 5's $11,181.87. Vals AI ReverseEngBench has agents recover the behaviour of real binaries, and GPT-6 fully solves 56.87 percent of them against 30.53 for GPT-5.6 Sol. Bug Hunt Bench is a new board of 60 entries with 105 defects planted in two production repositories. GPT-6 at max effort fixes 48 of them and Claude Fable 5.1 at max effort fixes 43. Claude Opus 5 leads OSWorld 2.0, 108 desktop workflows judged on full completion, at 34.72 percent with GPT-5.6 Sol at 28.1. Best measured model is Claude Mythos 5 at 2131.
Elsewhere. Incompressible Knowledge Probes is new. Its 133 entries answer closed-book factual questions across seven obscurity tiers running from common facts to obscure Wikidata entities, and Gemini 3 Flash (Preview) at medium effort leads at 86.19 percent with Gemini 3.1 Pro (Preview) second at 85.60. ChessBench has 61 entries playing rated games against fixed reference opponents, led by Gemini 3.8 Flash at 75.9 percent. A human reference row now sits atop the Last Translation Benchmark at 99.89 percent. The best model there is GPT-6 Pro at 45.33.
