CoT steerability drops with length across GPT-6 and Fable 5.1, possibly industry-wide

On September 4, blogger 1a3orn published a thread comparing chain-of-thought (CoT) steerability data for OpenAI's GPT-6 (Astra) and Ant's Fable 5.1, concluding that steerability drops sharply as CoT length grows—and this isn't unique to OpenAI but likely an industry-wide issue.

Confirmed

At around 2k tokens of prompt length, GPT-6 (Astra) steerability is roughly 50%, falling below 10% at 10k tokens (the original report's charts use a logarithmic scale)
Fable 5.1 shows 50%-65% steerability at around 2k tokens, and also drops below 10% at 10k tokens
1a3orn notes that Astra has a firmer grip on its chain of thought than earlier models (i.e., lower steerability), though the magnitude needs to be judged from the charts
Compared against Claude, Fable 5.1's steerability is higher than all models except Mythos preview, and it also degrades with length

Unconfirmed

1a3orn himself points out that the original report's charts are rough and hard to read, so the true steerability gap between the two models' CoT awaits clearer data
He notes it's also plausible that Astra is "better" (less steerable) at certain CoT lengths, suggesting room for interpretation of the exact curve shapes

Why It Matters

1a3orn argues that degrading CoT steerability with length is unlikely to be an OpenAI-specific looped transformation problem, hinting at a shared industry weakness in supervising long chains of thought
He questions whether the public has been unfairly harsher on OpenAI than on Anthropic, offering a counterweight to the controversy surrounding GPT-6

2026-09-04 ~ 2026-09-04 · 6 related posts

Primary sources

GPT 6 chain-of-thought controlability falls below 10% at 10k tokens, analysis of reports finds — 1a3orn · 2026-09-04 06:54 (UTC+08:00)
Cross-model comparison suggests CoT controllability decline isn't OpenAI-specific — 1a3orn · 2026-09-04 07:04 (UTC+08:00)

[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