Skip to content
ThreadOut
Jay Chooi@chooi_jeq · Sep 5, 2026

GPT-6 Astra scored 95% on a robot control task, up from Fable 5.1's 40%, with 6.2x fewer output tokens at 2.3x lower…

10 tweets1 min read♥ 2.5Koriginal

Summary

GPT-6 Astra significantly outperforms Fable 5.1 on robot arm control tasks, achieving 95% success on standard tasks with 6.2x fewer tokens and 2.3x lower cost. Researchers suggest LLMs could control robot arms in real-time within 2-4 years and call for the community to study implications of general-purpose robots arriving soon.

Summarized by ThreadOut AI from the full thread. May miss nuance — read the thread below.

  1. #1

    GPT-6 Astra scored 95% on a robot control task, up from Fable 5.1's 40%, with 6.2x fewer output tokens at 2.3x lower cost. 🧵

  2. #2

    On harder tasks involving precision, Astra has the same success rate as Fable 5.1 (2/20), but uses 3.9x fewer output tokens and is 1.6x cheaper.

  3. #3

    We let Astra and Fable control the arms via specifying the end-effector pose and passing it through an automatic IK solver.

    Jay Chooi@chooi_jeq · Sep 4, 2026

    Fable 5.1 POV. Fable writes an end-effector pose and an IK solver turns it into joint commands. LLMs can also drive the joints directly with no solver as in our previous results.

  4. #4

    If trends hold, LLMs could control robot arms in real time as soon as the end of this year, or by 2029 at the latest.

    Jay Chooi@chooi_jeq · Aug 20, 2026

    LLM token output speed increases by 2-7x per year, with Fable-class models doubling every month (month!). If trends continue, LLMs could meaningfully control robots in real time by end of the year, or by 2029 at the latest.

  5. #5

    We urge the research community to look into the implications if general-purpose robots arrive within the next two years. SOTA LLMs are showing surprising generalizability to robotics and their latency is improving rapidly.

  6. #6

    We release all 120 traces (transcripts, camera recordings, and trajectories) here: openai.robocurve.org/gpt-6-astra/

  7. #7

    We use Inspect Robots to evaluate VLAs, WAMs, and LLMs like Astra on robots. Inspect Robots is fully open-source: inspect-robots.org

  8. #8

    Robocurve (@robocurve) is a Public Benefit Corporation measuring frontier robotics capabilities and reporting them to the public. Learn more: robocurve.org

  9. #9

    We're hiring for Member of Technical Staff ($170k-$300k). $5k for successful referrals. Apply or refer here: robocurve.org/careers/

  10. #10

    Shout out to @achumenon_ for heroically leading this rapid-response eval!