Of 2,993 performance claims in the tensor.news record, just 559 have been independently reproduced. The other 2,432 — 81.3% — are third-party, self-reported or unverified, and two are contradicted outright. The evidence base for frontier models remains largely unconfirmed four months on. The July headline of 29% was 573 reproductions out of 1,943 claims; the ledger has since grown by 1,050 claims while the reproduced count fell by 14.
Anthropic's Claude Fable 5.2 tops the frontier desk with 14 benchmark leads — ARC-AGI, Chess Puzzles, DTBench, FrontierMath-Tier-4-v2-Private, FrontierMath-Tiers-1-3-v2-Private, FrontierSWE, and eight more. Not one of its 19 claims is independently reproduced. Every one is unverified.
That splits the top ten cleanly. Five models carry independently reproduced status — Claude Fable 5, GPT-5, GPT-5.5, Claude Opus 4.6, Gemini 3 Pro — and five carry none: Claude Fable 5.2, GPT-6 Astra, Gemini 4 Carbon, Claude Opus 5.5, Claude Fable 5.1. The July claim that no frontier model in the top ten has an unverified lead no longer holds; half the desk does.
The split follows release date. The five verified models all shipped in 2025 Q3 through 2026 Q2, and their reproduced scores carry Epoch AI's name. The five unverified ones are all 2026 Q3 and Q4 releases, and their scores are the ones the September export left unsourced. Claude Fable 5.2, GPT-6 Astra, Gemini 4 Carbon, Claude Opus 5.5 and Claude Fable 5.1 have 107 claims between them and not one reproduction.
Two benchmarks now hold a perfect integrity score of 100: Codeforces rating and Mystery Game Puzzles. FrontierMath-Tier-4-v2-Private has no headroom at all — the top score is the ceiling — and now sits at 99 with Chess Puzzles, FrontierMath-Tiers-1-3-v2-Private, Furniture Assembly, MirrorCode and Remote Labor Index; three more follow at 98: FrontierMath-2025-02-28-Private, LiveCodeBench and SimpleBench. Thirty-six of the 63 tracked benchmarks now grade A.
Codeforces rating earns its 100 without a ceiling to measure against: it scores on an Elo-style rating with no fixed maximum, so there is no headroom for the leaders to crowd into, and the record says so rather than inventing a bound.
At the bottom, ten benchmarks grade C or D with contamination as the weakest axis: LAMBADA (54), ANLI (59), TriviaQA (63), OpenBookQA (65), MMLU, ARC AI2 and PIQA (66), Winogrande and HellaSwag (67), GSM8K (68). Sixteen benchmarks score zero on contamination, and they hold 734 model scores between them. MMLU and ARC-AGI are both on that list.
One July benchmark is gone. HLE had a perfect integrity score and appeared in the top five; the current record carries no HLE rows at all, and the 63 tracked benchmarks are a different set from the 53 in the July export.
Among the 73 priced models that also carry benchmark scores, 26 prices are corroborated across sources, 30 are disputed, and 17 rest on a single source. More listed prices are contested than confirmed.
The cheapest are contested. GPT-5 nano at $0.1375 per million tokens is disputed, as are Mistral NeMo ($0.164) and Qwen3-235B-A22B ($0.30). GPT-4.1 nano ($0.175) and GPT-4o mini ($0.262) are corroborated. At the top, GPT-5.5 Pro and GPT-5.4 Pro share $67.50, both corroborated; GPT-5.2 Pro at $57.75 is disputed. The spread from $0.1375 to $67.50 is 491-fold.