AI
“OpenAI lançou o 'GPT Astra 6', um agente de IA que assume controle do computador para executar tarefas autonomamente (navegar em interfaces, preencher formulários, abrir softwares como Blender, construir e testar sites, editar contratos), atingindo 72,6% no benchmark OSWorld 2.0 (o melhor resultado já visto em uso de computador), 47%…”
Plain restatementOpenAI released a flagship model with computer-use capability that scored 72.6% on OSWorld 2.0, the highest computer-use result recorded to date, completed tasks in 47% less time than the prior model, and supports a 1 million token context window.
Distortion codes this site does not recognise yet: benchmark_cherry_picking, harness_mismatch, demo_to_product_conflation, capability_extrapolation, scale_conflation. Not collectible until the field guide has an entry.
The model is real, but the post oversells it. OpenAI did launch a computer-use model on September 3, 2026, though its name is GPT-6 Astra, not "GPT Astra 6," and the 72.6% score, the 47% time saving and the roughly one million token context window all come straight from OpenAI's own launch announcement. What the post leaves out is what those numbers actually measure: OpenAI ran the test itself, on an offline subset of the benchmark, using the partial-credit metric rather than the benchmark's main pass-or-fail metric, on which no system anywhere has topped about 32%. The claim that this is the best computer-use result ever recorded is OpenAI's own marketing line, and it is contested, since Anthropic reports a higher partial score for Claude Fable 5.1 and one tracked leaderboard put Claude ahead in early September. The two results were measured on different versions of the test, so neither side can currently claim the top spot honestly. The impressive demos, including the Blender file and the website built and tested without human help, come from OpenAI's own launch page, and OpenAI's fine print states the clips shown are edited excerpts, with no independent reproduction found. Treat the model as a genuine and significant step forward, and treat the "best ever, fully autonomous" framing as advertising rather than verified fact.
OpenAI [drifted from the evidence:] lançou o 'GPT Astra 6', um agente de IA que assume controle do computador para executar tarefas autonomamente (navegar em interfaces, preencher formulários, abrir softwares como Blender, construir e testar sites, editar contratos), atingindo 72,6% [drifted from the evidence:] no benchmark OSWorld 2.0 [drifted from the evidence:] (o melhor resultado já visto em uso de computador), 47% [drifted from the evidence:] menos tempo por tarefa que o modelo anterior, e suportando 1 [drifted from the evidence:] milhão de tokens de contexto.
OpenAI [added by the neutral restatement:] released a flagship model with computer-use capability that scored 72.6% [added by the neutral restatement:] on OSWorld 2.0, [added by the neutral restatement:] the highest computer-use result recorded to date, completed tasks in 47% [added by the neutral restatement:] less time than the prior model, and supports a 1 [added by the neutral restatement:] million token context window.
Red-tinted words in the claim drifted from the evidence. Green-tinted words are what a neutral restatement needs.
The trace / claim to source
- OpenAI did launch a flagship computer-use model on September 3, 2026, and it is real and shipping.
- 72.6% on OSWorld 2.0 is the correct figure as published by OpenAI. It is not invented.
- The 47% time reduction is correct as published: roughly 40 minutes per task for Astra against roughly 75 for GPT-5.6 Sol.
- The 40-minute average figure quoted in the post's image is correct.
- The context window claim is essentially correct. The actual figure is 1,050,000 tokens.
- The general capability description is directionally correct. OpenAI does describe Astra as operating interfaces, filling forms, updating records, and driving engineering software, and does present Blender, website and legal-document demonstrations.
- Omitted qualifier: the post gives 72.6% as a bare OSWorld 2.0 score. OpenAI's number is on an offline subset, on the partial-score metric, from a latency simulation, at maximum effort. The benchmark's own primary metric is binary completion, where the best tracked system across the whole board is 32.0%. A reader takes 72.6% to mean the agent finishes roughly three of four real tasks. On the benchmark's headline metric, no system is close to that.
- Marketing as evidence: every number in the post is OpenAI's own, run by OpenAI, in OpenAI's environment. The post presents them as measured facts with no attribution to who ran them. No independent runner result for Astra on OSWorld 2.0 was found.
- Benchmark cherry picking: "o melhor resultado já visto em uso de computador" is stated as settled. Anthropic reports 77.9% partial for Claude Fable 5.1 on OSWorld 2.0, and one tracked aggregator listed Fable 5.1 as leading on September 4, 2026. The two figures are genuinely not comparable across task releases, which is precisely why "best ever" is not a claim the evidence currently supports in either direction.
- Demo to product conflation: "construiu e testou um site inteiro, do código ao QA visual, sem intervenção humana" is presented as verified autonomous performance. OpenAI's own footnote states the displayed clips are edited excerpts, and at least one analyst documented a run listed at 2 minutes 54 seconds shown as a 15-second condensed clip. Edited excerpt is not the same as unattended completion.
- Capability extrapolation: the post moves from benchmark and demo results to "um agente trabalhando na sua tela enquanto você faz outra coisa" as a present workflow reality. Even taking OpenAI's own figure at face value, roughly one in four benchmark tasks fails, and the benchmark is a research environment, not a production desktop.
- Scale conflation (minor): the model is GPT-6 Astra. "GPT Astra 6" is not the product name. It does not change which model is meant, but it signals the post was not written from the primary source.
- Omitted qualifier (access): the post frames the model as simply launched. Initial access was gated to a limited set of organizations because Astra is the first OpenAI model classified at the Critical cybersecurity threshold under its Preparedness Framework, a fact absent from a post aimed at a general business audience.
- GPT-6 Astra's binary-completion score on OSWorld 2.0 was not found in any source. Without it, the 72.6% cannot be placed against the board's primary metric.
- Whether Astra genuinely leads computer use cannot be resolved today, because OpenAI and Anthropic scored on different OSWorld 2.0 task releases and neither result has an independent reproduction on a common harness.
- I read the OpenAI launch post, the Snorkel leaderboard and the Steel.dev board through search-index excerpts rather than loading the pages in full, so I could not read the complete benchmark tables or confirm the boards' current state beyond the excerpted text and their stated update dates.
- The specific demos named in the post, notably the Blender 3D print file and the end-to-end site build with visual QA, could not be matched one-to-one to a labelled OpenAI demonstration with published conditions. Independent reproduction of any of them was not found.
- Whether the phased rollout has since reached all Plus and Business users was not verified beyond OpenAI's September statements.
The underlying product is real, and every headline number in the post traces to one artifact: OpenAI's own launch post of September 3, 2026. The model's actual name is GPT-6 Astra, not "GPT Astra 6." OpenAI's launch post states that in latency simulations on OSWorld 2.0, Astra reached 72.6% at roughly 40 minutes per task, against 65.7% at roughly 75 minutes for GPT-5.6 Sol. That is the source of both the 72.6% figure and the 47% time reduction. OpenAI describes Astra as state of the art on computer use, browsing, software engineering, cybersecurity and professional work. Three qualifiers sit on that number in OpenAI's own material and in launch coverage. VentureBeat and Atlas Cloud both describe the score as coming from an offline subset of OSWorld 2.0, and Atlas Cloud specifies OpenAI's "offline partial-score set." OpenAI's own footnote block explains that OSWorld V2-Offline is a subset that runs without internet access, and adds a separate footnote that the displayed demonstration clips are edited excerpts. The metric distinction is material. The OSWorld 2.0 authors score two ways: binary completion, meaning all checkpoints passed, and a partial score, meaning the fraction of checkpoints reached. The paper reports that under the primary binary-completion metric at a 500-step budget, the best system it tested completed only 20.6% of tasks at a 54.8% partial score. The Snorkel AI leaderboard, run by a benchmark co-author, shows Claude Opus 5 leading at 31.43% binary and 68.31% partial, with GPT-5.6 Sol at 27.34% binary and 62.72% partial. GPT-6 Astra does not appear in the excerpt of that co-author board. No binary-completion figure for Astra could be found anywhere. On the superlative, the picture is contested rather than settled. Anthropic reports Claude Fable 5.1 at 77.9% partial and 41.7% strict on OSWorld 2.0, higher than Astra's 72.6%, and the Steel.dev tracked board listed Fable 5.1 as the leading tracked setup at 77.9% as of September 4, 2026. Those numbers are not directly comparable: DataCamp notes Anthropic scored on the benchmark authors' August 2026 task release and states the numbers are not comparable to previously published OSWorld 2.0 results, while OpenAI's footnote says its Claude comparison figures use official settings rather than the modified tasks and grading from the Fable 5.1 system card. Aggregators split accordingly, with BenchLM.ai placing Astra first at 72.6% and Steel.dev placing Fable 5.1 first at 77.9%. Context window: OpenRouter, eesel and Layer3Labs all list 1,050,000 tokens, so "1 million" is a fair rounding. Release status: OpenAI's official account said the model was rolling out to a limited set of organizations with broader availability over coming days, and a later official post confirmed availability to Pro, Enterprise and Business Premium users in ChatGPT Work and Codex plus the API. CNBC reports the phased rollout began with participants in OpenAI's cybersecurity program, gating tied to Astra being the first OpenAI model to reach the Critical threshold under its Preparedness Framework.
Complete reasoning
The reply is formatted for pasting into the thread where the claim is circulating.
Compact share page: ai.trueseeker.com/s/52da3bb51865/8SjZeKpnCrQ9Evlep6aw7-N98zE
Ask this case
Answers come only from the case file above; nothing is added.
Is GPT-6 Astra a real product from OpenAI?
Yes. OpenAI launched a flagship computer-use model on September 3, 2026, and it is real and shipping. Its actual name is GPT-6 Astra, not "GPT Astra 6" as the claim states.
Is the 72.6% OSWorld 2.0 score accurate?
The number itself is accurate as published by OpenAI, but it comes from an offline subset of the benchmark using the partial-score metric, not the benchmark's main binary pass-or-fail metric. On that primary metric, the best tracked system anywhere scores only about 32%, so 72.6% does not mean the agent completes three out of four real tasks.
Is it true this is the best computer-use result ever recorded?
That claim is contested, not settled. Anthropic reports a higher partial score of 77.9% for Claude Fable 5.1, and one tracked leaderboard showed Claude ahead in early September, though the two results were measured on different versions of the test so neither side can currently claim the top spot honestly.
Did the AI really build and test a website with no human help?
That demo comes from OpenAI's own launch page, but OpenAI's own fine print says the clips shown are edited excerpts, and no independent reproduction was found. An edited excerpt is not the same as a verified unattended completion.
Are the 47% time savings and 1 million token context window accurate?
Yes, both are correct as published by OpenAI: roughly 40 minutes per task versus about 75 minutes for the prior model, and a context window of about 1,050,000 tokens, which rounds fairly to 1 million.