Benchmark scores drop from 89% to 19% on new evals — how benchmaxxing breaks leaderboard trust
airesearch12 · x · 2026-09-11
A post on "benchmaxxing" claims Google and Meta's models look nearly as good as OpenAI and Anthropic's top models on standard benchmarks (89.4% and 88.8%), yet drop to 19.1% and 33.3% the moment a fresh benchmark drops, while GPT and Fable models hold steady at 56-58%.
The takeaway: labs optimizing around public benchmarks instead of generalizing make leaderboard scores increasingly untrustworthy. The author says they've found a way to measure benchmaxxing and may release a public board scoring all models.
Caveat: the model names cited (Gemini 3.8 Flash, GPT-6 Astra, etc.) are not real released models — treat this as illustrative or satirical, not factual reporting.
More from Models
- AI Sextet offers 6 models free and unlimited for 14 days, including DeepSeek and Qwen — airesearch12 · 2026-09-11
- BullshitBench update: GPT-6-Astra beats all prior OpenAI models but still trails Anthropic — scaling01 · 2026-09-11
- Astra Scores 83% on GauntletBench, First Computer-Use Agent to Beat Human Baseline — ducha_aiki · 2026-09-11
- Kimi K2.8 Preview rolls out: near-K3 coding performance, 1M context for all tiers — teortaxesTex · 2026-09-11
- Looking for a classifier of software engineering task shapes to pick models per task — StewartalsopIII · 2026-09-11
- DeepSeek V4 Pro API to continue after Sept 2026, billing unchanged — teortaxesTex · 2026-09-11