A story we follow
OpenAI releases GPT-6 Astra
Tracks Astra's release, demonstrated capabilities, benchmark performance, computer-use design and deployment restrictions.
A story page follows one specific event across podcast discussions, with an overview and a timeline of what changed. It updates when new episodes discuss the event, and you can get those updates by email or push. How our formats work
By email or push, when we publish an update to this story. It does not subscribe you to other stories.
Overview
OpenAI introduced GPT-6 Astra with staged access, claims that it saturated benchmarks, and a reported Critical cybersecurity classification. Early discussions focused on its computer-use and token efficiency, agent simulations of a city, and faster internal development at OpenAI.
Much of the later discussion has been about physical control. Robocurve benchmarks on robotic arms showed high completion rates on basic manipulation, such as moving blocks and picking up cups, at lower cost than rival models. Precision insertion remained hard. In the DrivingBench test, general-purpose models issued bounded commands to drive a 2022 Toyota Corolla through a cone course using openpilot. Astra was the only model to finish, on its second of up to three attempts. The authors note that the harnesses differed, steering range was limited and no sessions were repeated.
People read these results in different ways. In September, Moonshots hosts argued that generalist models make specialized robotics models obsolete. Dave Blundin countered that a second-try finish is not road-ready and that compute use was not reported. On Training Data on 20 October, Sequoia's Sonya Huang asked whether Astra's robot results mean the AGI labs will "clobber" pure-play robotics companies. Mecka founder Josh Gao, whose company sells robot training data, said the results were expected, because robotics data is now being added to the pre-training mixes of large foundation models. He also said specialized robotics models might still reach 99.99 percent reliability on particular tasks where generalist models are only very good.
Open questions remain about deployment restrictions, compute costs, and how far generalist models' benchmark results carry over to reliable physical control.
What changed
Dates show when each podcast discussion was published.
-
Interpretation
On Training Data, Sonya Huang said Astra was generalizing well to robot tasks and asked whether AGI labs will steamroll pure-play robotics companies. Josh Gao of Mecka, a robot-data supplier, called the results "quite expected" because robotics data is being added to the pre-training mixes of foundation models. He argued that specialized robotics models might still reach four-nines reliability on specific tasks. This is opinion; neither speaker cited a specific evaluation.
-
New information
Moonshots discussed DrivingBench, in which Astra alone completed a real-car cone course, finishing on its second attempt. Alex Wissner-Gross said generalist models make specialized robotics obsolete. Blundin said a second-try finish is not road-ready and that the compute cost was left out. The developers note their own limits: differing harnesses and no repeated sessions.
-
New information
The panel highlights evaluations of Astra plugged into robotic arms for embodied physical manipulation. Citing Robocurve benchmark results, panelists discuss the model achieving near total completion rates on basic tasks like moving blocks and picking up cups at lower inference costs than rival frontier models, while noting that precision manipulation remains an open challenge.
-
New information
The panel describes Schumer using Astra to build Manhattan and simulated agents that communicate and create another simulation. It also quotes a Codex engineering lead saying Astra accelerated some plans from mid-next-year delivery to Dev Day.
-
New information
The panel interprets Astra as a computer-use and token-efficiency advance, contrasting benchmark saturation with weaker results on broader capability and cost comparisons. It also discusses restricted deployment following the reported Critical cybersecurity assessment.
Podcast discussions
- Training Data Inside the Robot Data Factory: Mecka's Josh Gao on Sensorizing Humans to Train RobotsOct 20, 2026
- Moonshots with Peter Diamandis Why Jensen and Zuck think the doomers are wrong (plus AI get’s a rebrand) | #294 MOONSHOTS LiveSep 28, 2026
- Moonshots with Peter Diamandis Should we slow down AI progress? | MOONSHOTS #288Sep 28, 2026
- Moonshots with Peter Diamandis Why Jensen Huang Believes We’ve Reached AGI and Inside OpenAI’s German Website Hijack | #287Sep 9, 2026
- Moonshots with Peter Diamandis GPT-6 Astra Saturates ARC-AGI-3, Tesla's $30K Cybercab Floods Austin, Anthropic Proves Fermat's Last Theorem | EP #286Sep 5, 2026
Sources
- 01
- 02
- 03
- 04
- 05
- 06
- 07
- 08
- 09
- 10
- 11
- 12
Version history
-
Oct 30, 2026 · Version 5
On Training Data, Sonya Huang said Astra was generalizing well to robot tasks and asked whether AGI labs will steamroll pure-play robotics companies. Josh Gao of Mecka, a robot-data supplier, called the results "quite expected" because robotics data is being added to the pre-training mixes of foundation models. He argued that specialized robotics models might still reach four-nines reliability on specific tasks. This is opinion; neither speaker cited a specific evaluation.
-
Sep 27, 2026 · Version 4
Moonshots discussed DrivingBench, in which Astra alone completed a real-car cone course, finishing on its second attempt. Alex Wissner-Gross said generalist models make specialized robotics obsolete. Blundin said a second-try finish is not road-ready and that the compute cost was left out. The developers note their own limits: differing harnesses and no repeated sessions.
-
Sep 15, 2026 · Version 3
The panel highlights evaluations of Astra plugged into robotic arms for embodied physical manipulation. Citing Robocurve benchmark results, panelists discuss the model achieving near total completion rates on basic tasks like moving blocks and picking up cups at lower inference costs than rival frontier models, while noting that precision manipulation remains an open challenge.
-
Sep 15, 2026 · Version 2
The panel describes Schumer using Astra to build Manhattan and simulated agents that communicate and create another simulation. It also quotes a Codex engineering lead saying Astra accelerated some plans from mid-next-year delivery to Dev Day.