Open benchmarking challenge says Opus 5’s ARC-AGI-3 leap doesn’t carry over
rbhar90 · x · 2026-07-25
The post argues for more open benchmarking, saying it no longer trusts strong reasoning claims from labs because too much money is now tied to the narrative.
It quotes a comparison around Opus 5 on ARC-AGI-3 versus the authors’ held-out Witness suite:
- Opus 5 reportedly scores 30% on ARC-AGI-3, far ahead of prior models.
- On Witness, the jump does not transfer: Opus 5 lands at 43.4 ± 3.2, essentially tied with kimi-k3 (42.8 ± 1.9) and Fable-5 (43.8 ± 9.7).
- It is better than Opus 4.8 (34.8), but the authors say this is not a generational leap.
- The traces suggest Opus 5 already “knows” this class of puzzle: it states hidden rules before acting and then plays a byte-identical optimal solution in all 5 seeds at temperature 1.0, with zero exploration.
The broader point is that benchmark scores can overstate reasoning gains if the evaluation set and the real task distribution diverge.
Related event: Opus 5's High ARC-AGI-3 Score Sparks Cheating and Overfitting Controversy(11 posts)→
More from Models
- 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
- DeepSeek V4.1 Flash Hits 98% of GPT-6 Astra's Score at 1.4% of the Cost in Third-Party Benchmark — ayushtweetshere · 2026-09-11
- TheZvi Polls: Has Your Coding Model Choice Changed Since Fable 5.1 and Astra? — TheZvi · 2026-09-11
- antirez Weighs In on Anthropic Banning Minors From Using Claude — antirez · 2026-09-11
- Engram's random reads don't suit SSDs; CPU-memory over NVLink could serve all 72 GPUs — bookwormengr · 2026-09-11