Revision History

Update: add a "Robot control" subsection under Capabilities and benchmarks on third-party robot-control evaluations of GPT-6 Astra (Robocurve reports and RoboDojo-RC Tier 1 leaderboard, Artificial Analysis AA-Robotics v0.1, Embodied First Exam, SenseNova technical report), with references 130-138.

Update: noted that OpenAI's October 23, 2026 containment post names an upcoming GPT-6 Astra Cyber model among its red-teaming agents.

Update: added the developer's official video.

Auto-saved before update

Update: added The Information's October 2026 report that DeepSeek and Moonshot AI relied on Astra outputs to train their latest models, with a link to distillation attack

Update: added GPT-Rosalind-6 (October 30, 2026), a trusted-access life sciences model that uses GPT-6 Astra's weights

Update: add Unitree founder Wang Xingxing's October 22, 2026 remarks on robot experiments with GPT-6 Astra

Correction: Astra is no longer the only Ultrafast model (GPT-6.1 Sol was added October 8, 2026); updated Ultrafast rate limits for OpenAI's October 6 three-tier system and noted that Astra's Ultrafast mode lacks EU data residency.

Update: link the new CritPt article

Verification fixes to the GPT-6 Pro paragraph: naming hedge and full release-note quote

Add GPT-6 Pro naming in ChatGPT and the October 16, 2026 GPT-6 Pro upgrade (faster responses, plugins/tools, parallel agents); refs 117-120

Improve internal links (2026-10-13)

Correction: the page did not mention Ultrafast, the premium speed tier OpenAI added for GPT-6 Astra on September 29, 2026 and the only model it covers. Added a sourced paragraph with the API prices, rate limits, speed claims and ChatGPT usage multipliers, plus five sources.

Update: added the September 29 "Appendix: dots" system card change-log entry, the GPT-6.1 Sol launch, the reported cancellation of GPT-6.1 Astra, and the 68.1% Terminal-Bench Science figure OpenAI later gave for Astra

Update: OpenAI's 21 Sep 2026 Advisory Group announcement and its claim of 100+ open problems resolved by its internal model

Minor: removed speculation about why OpenAI's three honeypot figures differ; the card's stacked Figure 9 bar separates successful attacks, attempted attacks and legitimate exploits

Minor: note that the card's own Figure 9 labels Sol's honeypot bar 56.0% while its text says 55.4%; deception figures located in the card's Figures 10 and 11

Minor: the launch post still shows the pre-revision HealthBench 63.4% after the system card was revised

Correction: GPT-6 Sol and Luna released Sept 22 (no-variants sentences dated); system card HealthBench values for Astra revised Sept 22; updated alignment evals noted

Verification fixes: BenchCAD sample-size context; OSWorld footnote quote updated (OpenAI reworded it after launch; launch-day wording noted)

Add OpenAI's own BenchCAD description (3D reconstruction from multi-view renders via CAD code) and cost note, The New Stack's Vision2Code-subset detail, and the 'visual judgment ... renderings' line; link inverse_graphics

Auto-saved before update

Added the launch post's mid-task steering description and the September 4 OpenAI Developers video post to the Codex section

Link Claude Fable 5.1 and Claude Mythos 5.1 to their new pages

Correction: internal ExploitBench (June-August 2026) Sol figure is 11.5% (launch-day table cell read 5.5% against the post's own chart); Fast-mode speed and Bedrock wording changes on the live post noted

No summary provided