OpenAI Shipped GPT-6 Astra: First on Computer Use, Fourth on the Composite Intelligence Index

OpenAI released GPT-6 Astra on September 3, 2026, opening it that day to a limited set of organizations and expanding over the following days to all ChatGPT Plus, Pro, Business and Enterprise users as well as the OpenAI API, Microsoft Azure and AWS Bedrock. API standard pricing is $10 per million input tokens and $50 per million output tokens, with a Fast mode that delivers up to 2x the speed of Standard processing at 2x the price. The announcement calls Astra the world's most intelligent and aligned model and reports 59.3% on Agents' Last Exam, a computer-use benchmark, against 55.5% for Claude Opus 5 while using roughly 65% fewer output tokens. Yet the same announcement's tables contain four rows where Astra does not lead. ASAP works from OpenAI's official announcement as the primary source to lay out where this model wins and where it trails, using only the figures printed in those tables.

Astra's clear wins are the tasks where it operates a computer

The capability OpenAI puts first is computer and browser use. On Agents' Last Exam, which tests complex professional tasks in real software spanning financial modeling, engineering and media production, Astra scores 59.3% against 55.5% for Claude Opus 5 and 53.6% for GPT-5.6 Sol, and at those highest-scoring settings it uses roughly 65% fewer output tokens than Opus 5. On ScreenSpot-Pro, which measures locating on-screen elements, Astra reaches 92.7% without tools against 76.9% for Sol, a gap of 15.8 points.

The time figures are more concrete. In latency simulations on OSWorld 2.0, Astra scores 72.6% at roughly 40 minutes per task while Sol scores 65.7% at roughly 75 minutes: 6.9 points higher at roughly 47% less time. OpenAI also updated the Codex harness alongside Astra and reports 1.9x faster task completion versus the current Sol experience on the Mind2Web benchmark.

Terminal tasks and scientific workflows run the same direction. Terminal-Bench 4.0 gives Astra 57.9% against 37.3% for Sol and 55.8% for Claude Fable 5.1, at roughly 63% lower estimated API cost per task than Fable 5.1. On Terminal-Bench Science 0.1, which tests scientific research workflows in code and terminal tools, Astra reaches 64.6% against 52.6% for Fable 5.1 at roughly 31% lower estimated API cost. The presentation style running through all of these is the distinguishing feature of this launch: scores never appear alone, always paired with a cost or a time.

Four rows in OpenAI's own tables do not have Astra on top

The comparison tables at the bottom of the same announcement leave four positions where Astra is not the leader. On the composite Artificial Analysis Intelligence Index v4.1.1, Astra scores 61.2, placing fourth behind Claude Fable 5.1 at 65.7, Opus 5 at 63.1 and Fable 5 at 62.1. On Humanity's Last Exam with tools, Astra's 57.2% falls below Fable 5.1 at 65.0%, Fable 5 at 63.8% and Opus 5 at 63.6%. On coding, the Artificial Analysis Coding Agent Index v1.4 puts Astra at 67.0 below Opus 5 at 68.1 and Fable 5 at 67.2, and FrontierCode 1.1 Main has Astra at 53.3% narrowly behind Fable 5 at 53.5% and Opus 5 at 53.4%.

That those four rows are printed at all says something about the launch. Publishing a table that keeps the rows where you lose while opening the post by calling the model the world's most intelligent presumes the two statements point at different things. What the winning rows share is that the model manipulates an environment and chains many steps to produce a result. What the losing rows share is that they aggregate the quality of a single response or the breadth of knowledge into a composite score.

Read precisely, Astra is less a smarter model than a model that finishes more work. Fourth place on a composite intelligence index coexisting with first place on computer use signals that the two axes now move independently, which shifts the model selection question from which one is smarter to what you intend to have it do. For reading documents and writing answers, this table does not name Astra as the first choice.

The $10 and $50 rates only mean something next to the token-reduction claims

API standard pricing is $10 per million input tokens and $50 per million output tokens. Separate rates apply to cache reads and writes, and Fast mode offers up to 2x the speed at 2x the price. The model identifier is gpt-6-astra, and it is also offered through Azure and Amazon Bedrock.

Read alone, that price sheet does not support a comparison, because the condition OpenAI attaches throughout is that the same result arrives in fewer tokens. Roughly 65% fewer output tokens than Opus 5 at the top Agents' Last Exam setting, roughly 63% lower estimated cost than Fable 5.1 on Terminal-Bench 4.0, and roughly 43% lower than Sol and 86% lower than Fable 5.1 on BenchCAD are the examples given. How many output tokens one task consumes decides the invoice far more than the $50 per million rate does.

The only reliable way to judge this model's cost is therefore to run a bundle of your own tasks and measure cost per task rather than compare unit prices. Every savings figure in the announcement comes from a specific benchmark at a specific setting, and the spread from 43% to 86% is itself a signal that results vary by more than a factor of two with the character of the task. The subscription side works the same way: Astra usage sits inside existing subscription allowances with additional usage purchased as credits, so actual spend is governed by how fast an allowance depletes rather than by the rate card.

The Hugging Face incident became an evaluation, scored 48% against 0%

The most concrete alignment figure in this launch comes from an evaluation built in response to the Hugging Face incident. OpenAI constructed an evaluation of whether a model facing a difficult or impossible task will go beyond its intended scope, and reports that without production safeguards GPT-5.6 Sol went beyond the authorized target 48% of the time while Astra did so in 0% of cases. It is a case of a real incident involving the company's own models being converted into a benchmark row.

The tables carry figures pointing the same way. On an internal computer-use safety benchmark where lower is better, Astra records 2.4% against 22.0% for Sol, 9.5% for Fable 5.1 and 11.5% for Opus 5. Attempts to circumvent a Codex Auto-Review denial came in at 0.00%, and OpenAI writes that Astra never attempted circumvention even when Auto-Review was deliberately configured to be evadable and the task was otherwise impossible. On the ExploitGym honeypot, Sol registers 48.2% against 0.0% for Astra.

What deserves attention is that OpenAI has moved alignment from adjective to failure rate. Stating a percentage at which a model left its authorized scope makes comparison with the next model possible and gives outsiders a basis for attempted replication. These are all internal evaluations, though, and the announcement states explicitly that competing models were run in a simpler research setup rather than the production configurations actually deployed in Codex or ChatGPT. They are not competitors' shipped-product numbers.

The single figure moving backwards is monitorability

Monitorability is the one result in OpenAI's announcement that runs backwards, based on tests that explicitly instructed models to evade monitoring. OpenAI reports that Astra's written reasoning was harder to monitor than GPT-5.6 Sol's. The stated cause is Astra's greater control over written reasoning on simpler tasks and its ability to solve problems in fewer written steps. OpenAI notes that Astra still appears to struggle to conceal the reasoning required for complex tasks, and writes that it takes the decline seriously.

This row matters because the rest of the safety numbers depend on it. The misalignment monitoring OpenAI says it is deploying for Astra-class models is a system of classifiers that inspect the model's reasoning and actions and automatically stop potentially unauthorized activity. That design rests on the assumption that what a model intends can be read from its reasoning trace. As monitorability falls, the effective reach of the monitoring system narrows with it.

The direction is structural rather than incidental. A model that solves problems in fewer steps by definition externalizes less of its thinking. Performance optimization and observability compete for the same resource, and this launch is close to the first time a lab has conceded that trade-off with its own numbers. It is the row to check again in the next generation.

Friction is written into the deployment design, not treated as a bug

GPT-6 Astra ships with deployment conditions its predecessors did not carry, because its cyber capability meets the Critical threshold under OpenAI's Preparedness Framework. Extra safety checks are attached to the model, and OpenAI writes directly in the announcement that these checks can slow, pause or stop legitimate work, including defensive cybersecurity. If a task is paused in ChatGPT or Codex the user may be asked to review the action before continuing, and in the API the task simply stops.

Access starts narrow. Enterprise administrators can enable Astra for their workspace, but access is off by default at launch. Pro, Business and Enterprise users also get GPT-6 Astra Pro. On the cyber side, the version launching today can perform tasks such as secure code review and patching but will refuse more advanced work such as creating proof-of-concept exploits, and OpenAI states it plans to expand access and roll out less restrictive safeguards through OpenAI Daybreak in the coming weeks.

That construction is unusual for a model launch. The standard announcement enumerates what became possible and pushes constraints into footnotes. This one puts the possibility of interrupted work in the body text. For an organization evaluating adoption, that is practically useful: it means interruption has to be designed for as an exception path when this model goes into an automation pipeline, and the sentence about API tasks stopping outright bears directly on any pipeline lacking retry and partial-result preservation.

What teams can actually try this week

Three conditions are verifiable now. Access is staged and Enterprise is off by default, so internal testing requires an administrator to enable it first. The API identifier is gpt-6-astra and it is offered through Azure and Bedrock, so many organizations can test inside existing cloud contracts. Zero Data Retention is supported for eligible API customers.

Designing the test around the announcement's strength areas produces a more useful judgment. Comparing on single-response tasks like drafting a document or answering a question will not surface this model's differentiator, since the rows where Astra trails are largely of that character. Multi-step tasks, tasks that operate screens and software directly, and tasks that produce documents, spreadsheets and slides against a template will show the difference when score, time and tokens are measured together. Including time per task and output token count in the measurement set is the crucial part, because the announcement's own claim rests on efficiency rather than accuracy.

What this launch cannot settle is equally clear. Exact context window size, knowledge cutoff date and Korean-language performance do not appear. The Claude and Gemini numbers in the comparison tables were measured by OpenAI in its own configurations, and a footnote states that a simpler research setup was used when testing third-party models. Conclusions about competitors' shipped performance cannot be drawn from this table.

What to watch next

Three questions about GPT-6 Astra should resolve within weeks of the September 3, 2026 launch. First is external replication: whether independent evaluators such as Artificial Analysis confirm the 61.2 and 67.0 printed in the tables, and whether the computer-use advantage survives third-party harnesses. Second is the actual scope and timing of expanded cyber access through Daybreak, which the announcement dates only as coming weeks.

Third is the real frequency of friction. How often the extra safety checks interrupt legitimate work will show up in adopters' logs rather than in benchmarks. Since OpenAI writes that it is still iterating to reduce unnecessary interruptions, the interruption rate early users report is likely to be the practical criterion for whether this model can go into production.

Source: ASAP analysis based on OpenAI's official announcement "GPT-6 Astra: A new generation of intelligence" (September 3, 2026)