GPT-6 vs Fable 5.1: CoT Steerability Data Sparks Debate
On September 4, blogger 1a3orn posted a thread comparing chain-of-thought (CoT) steerability data for OpenAI's GPT-6 (Astra) and Ant's Fable 5.1, concluding that steerability declines sharply with CoT length and that this is likely an industry-wide problem rather than an OpenAI-specific one.
Confirmed
- GPT-6 (Astra) shows roughly 50% steerability at about 2k tokens of prompt length, falling below 10% at 10k tokens (the original chart uses log scale).
- Fable 5.1 shows 50%-65% steerability at about 2k tokens, also dropping below 10% at 10k tokens.
- 1a3orn notes Astra exerts stronger control over its CoT than prior models (i.e., lower steerability), though the magnitude requires reading the chart.
- Compared against Claude models, Fable 5.1's steerability is higher than all models except Mythos preview, with the same degradation trend.
- tenobrus (shared by MilesBrundage) similarly argues Fable 5.1 likely matches Astra on CoT steerability, with an example showing control achieved in very few characters, corroborating 1a3orn's cross-comparison.
Unconfirmed
- 1a3orn himself notes the original charts are crudely made and hard to read, so the true gap between the two models awaits cleaner data.
- He says it is plausible that Astra is actually "better" (lower steerability) at certain CoT lengths, leaving room for interpretation of the curve shapes.
Why it matters
- 1a3orn argues the length-dependent degradation is unlikely to be an OpenAI-specific looped transformation issue, pointing to a common industry weakness in monitoring long chains of thought.
- He questions why public criticism of OpenAI has been harsher than of Anthropic, offering a counterweight to the controversy around GPT-6.
2026-09-04 ~ 2026-09-04 · 7 related posts
Primary sources
- [source] GPT 6 chain-of-thought controlability falls below 10% at 10k tokens, analysis of reports finds — 1a3orn · 2026-09-04
- GPT-6's CoT controllability may be no worse than Fable 5.1's, argues researcher — 1a3orn · 2026-09-04
- Fable 5.1 shows under 10% CoT controllability at 10k tokens, ~50% at 2k — 1a3orn · 2026-09-04
- Fable 5.1 hits 50-65% controllability at ~2k-token CoT lengths — 1a3orn · 2026-09-04
- GPT-6 CoT controllability: ~50% at 2k tokens, under 10% at 10k — 1a3orn · 2026-09-04
- [source] Cross-model comparison suggests CoT controllability decline isn't OpenAI-specific — 1a3orn · 2026-09-04
- [source] Fable 5.1 likely matches Astra on CoT controllability, observers say — Miles_Brundage · 2026-09-04