Epoch AI rated Vending-Bench 2 ‘Flawed’ on Oct 21, 2026, citing exploitable environment bugs and unequal context across providers. Here’s the leaderboard and the review.

A 2026 re-run of the original slopsquatting research found 127 hallucinated package names that five unrelated frontier models — Claude,

A federal appeals court, not a benchmark, produced the most consequential documented finding in agentic commerce so far: in August

Anthropic’s own account of its 5.5-generation models contains a detail that undercuts the evaluations built around it: Opus 5.5 reportedly

Llama 4’s benchmark score came from an unreleased model, DeepSeek’s guardrails failed jailbreak tests, and Mistral hallucinates despite 256K context.

Epoch AI’s FrontierMath math benchmark got a 42%-of-problems rewrite in June 2026. Here’s what changed, why scores jumped, and where trackers disagree.

Ahrefs, Pew, Seer Interactive, and SparkToro data on how AI Overviews cut organic click-through rates in 2026 — and where the numbers disagree.

Grok’s 2026 record: rising hallucination rates, a UK ICO/Ofcom probe into image generation, and documented Grokipedia accuracy problems.

GDPval grades AI against real professional deliverables. The top score jumped from 38.8% to 74.1% in a year, but trackers disagree on who leads now.

Q3 2026 benchmark data: knowledge tests saturating, coding scores compressing, and agentic benchmarks fragmenting into versions that disagree by 20-60 points.