GPT 6 chain-of-thought controlability falls below 10% at 10k tokens, analysis of reports finds
1a3orn · x · 2026-09-04
Blogger 1a3orn compares CoT controlability charts from OpenAI's GPT 6 (Astra) and Ant's Fable 5.1 reports:
- GPT 6 drops to under 10% controlability at 10k tokens (log scale)
- Fable 5.1's report uses characters on the x-axis (tokenizer kept secret); eyeballing the chart gives roughly 50% controlability at 2k tokens (10k characters)
- Fable 5.1 outperforms everything except Mythos Preview, with the same reversal of controlability over time
- Net: no clear winner, but both frontier models degrade sharply as context lengthens
More from Models
- GPT-6 Astra 'crushes Tiny Computer Bench' by hacking an Amazon toy computer — OpenAIDevs · 2026-09-04
- GPT-6 Astra one-shots a 3D mockup tool nearly indistinguishable from reality — OpenAIDevs · 2026-09-04
- Apparent GPT-6 Astra demo shows model generating intricate peacock SVG in one shot — OpenAIDevs · 2026-09-04
- Astra reportedly trained on 100k GPUs at Stargate Texas in first $1B training run — ai · 2026-09-04
- Sean Taylor claims fast progress eradicating hallucinations; Andrew Ng: capabilities and safety can align — irinarish · 2026-09-04
- ValsAI says OpenAI's GPT 6 Astra has effectively saturated SRE-Bench reverse-engineering benchmark — sandersted · 2026-09-04