GPT-6 Astra reportedly scores 95% on robot control at ~43% of prior cost
A set of widely circulated numbers on X (Twitter) claims GPT-6 Astra scores 95% on a robot control task, versus 40% for the previous Fable 5.1, while using 6.2x fewer output tokens and costing 2.3x less (about 43% of the prior generation). The figures originated from a thread by @chooijeq and then propagated through a multi-level repost chain with identical numbers.
Confirmed
- @chooijeq's thread first gave these numbers: 95% on robot control (vs 40% for Fable 5.1), 6.2x fewer output tokens, 2.3x lower cost.
- @willcb, @AaronBergman18, @maxpaperclips, @chrisjpaxton, and @TinfoilTricorn (reposted by beffjezos) relayed the same figures, forming a clear propagation chain.
Unconfirmed
- The specific benchmark, evaluation setup, and source of the "robot control task" have no primary documentation; the 95% and cost figures exist only as secondhand thread claims without official papers or docs.
Why it matters
- If accurate, the numbers imply a general-purpose chat model substantially beating its predecessor on robot control while roughly halving inference cost, with significant implications for model selection in embodied AI.
- Community reactions split: @AaronBergman18 marveled that "robot control has been stuffed into a chat Transformer"; beffjezos declared "embodied AGI is near"; DJiafei quipped that GPT-6 should just call the MolmoAct2 checkpoint via API; Dorialexander used @maxpaperclips's repost to revisit a prior European debate about LLMs — underscoring the clear divide between general-model and specialized-model camps.
2026-09-05 ~ 2026-09-06 · 6 related posts
- Episode 1: GPT-6 Astra's 3D modeling demos flood in, community calls it SOTA(2026-09-04, 37 posts)
- Episode 2: Unverified 'GPT-6 Astra' Reportedly Generates Impressive 3D Models(2026-09-04, 2 posts)
- Episode 3: OpenAI Launches GPT-6 Astra and Declares the AGI Era(2026-09-04, 21 posts)
- Episode 4: GPT-6 Astra Sets New Records Across Benchmarks With More Intelligence Per Token(2026-09-05, 19 posts)
- Episode 5: GPT-6 Astra finds up to 176x code speedups in five minutes(2026-09-05, 2 posts)
- Episode 6: GPT-6 Astra demos flood social media, sparking both 'Cambrian explosion' hype and skepticism(2026-09-05, 9 posts)
- Episode 7: OpenAI's Astra Reportedly Trained on Over 100,000 GPUs(2026-09-05, 2 posts)
- Episode 8: GPT-6 Astra reportedly scores 95% on robot control at ~43% of prior cost(2026-09-05, 6 posts)
- Episode 9: GPT-6 Astra Reportedly Beats Gemini in Vision Benchmarks(2026-09-05, 3 posts)
- Episode 10: OpenAI Launches 24-Hour GPT-6 Astra Demo Challenge(2026-09-05, 4 posts)
- Episode 11: GPT-6-Astra tops MathArena leaderboard(2026-09-06, 2 posts)
- Episode 12: GPT-6 Astra Tops Code Arena at 1797, Fourth Leader Change in a Week(2026-09-06, 7 posts)
- Episode 13: Leaked Benchmarks Claim GPT-6 Astra Aces Enterprise Tasks(2026-09-06, 2 posts)
- Episode 14: GPT-6 Astra Rebuilds Maya Ruins 3D Model in Two Hours(2026-09-06, 3 posts)
- Episode 15: Developer Says OpenAI's Astra Is First Model to Make Real Progress on His Ultra-Complex Project(2026-09-06, 2 posts)
- Episode 16: Rumored GPT-6 "Astra" Beats Portal Autonomously, Unverified(2026-09-06, 7 posts)
- Episode 17: GPT-6 Astra saturates spatial reasoning benchmarks(2026-09-06, 4 posts)
Primary sources
- [source] GPT-6 Astra Scores 95% on Robot Control, 6.2x Fewer Tokens Than Fable 5.1's 40% — scaling01 · 2026-09-05
- GPT-6 hits 95% on robot control; joker asks if it just API-calls MolmoAct2 — chris_j_paxton · 2026-09-06
4 near-duplicate retellings: AaronBergman18 · willcb · max_paperclips · TinfoilTricorn