OpenAI releases GPT-6 Astra, a model built for coding, computer use and long-running agent tasks
Computer-use benchmarks and a million-token window show what agentic models are being optimised for, and where they still fail.
OpenAI has released GPT-6 Astra, a model aimed at coding, operating software directly, long-running agentic work, science and cybersecurity, InfoQ reported on 10 September. The model is designed to carry out multi-step tasks inside applications, with examples including filling forms, updating CRM records, building websites and analysing data.
It is available to ChatGPT Plus, Pro, Business and Enterprise subscribers, through the OpenAI API, and via Microsoft Azure and AWS Bedrock, though InfoQ notes the initial rollout is limited to selected organisations. It accepts up to one million tokens of input, and an experimental mechanism in Codex lets the model carry notes across context windows.
InfoQ cites 72.6 percent on OSWorld 2.0, a computer-use benchmark, against 65.7 percent for the earlier GPT-5.6 Sol; 57.9 percent on Terminal-Bench 4.0; 74.1 percent on DeepSWE v1.1; and 96.3 percent on long-context tasks between 512K and one million tokens. The reported hallucination rate is 4.2 percent, down from 12.2 percent.
GPT-6 Astra is described as the first OpenAI model classified at the "critical capability level" for cybersecurity under the company's Preparedness Framework, after testing found it could discover two previously unknown vulnerabilities and produce exploits against hardened systems. Nvidia chief executive Jensen Huang is quoted saying the model was trained on more than roughly 100,000 Grace Blackwell NVLink72 systems.
Why it matters
The headline benchmarks are no longer about answering questions but about operating a computer over many steps, where the industry expects the next commercial value. A million-token window plus a cross-window notes mechanism shows labs treating memory as a system design problem rather than something solved by a bigger context alone. The cybersecurity classification and the staged rollout point to capability thresholds starting to shape how, and to whom, frontier models are released.
The headline benchmarks are no longer about answering questions but about operating a computer over many steps, where the industry expects the next commercial value.
- OSWorld 2.0 (computer use)72.6%
- DeepSWE v1.1 (software engineering)74.1%
- Terminal-Bench 4.057.9%
- Long context, 512K to 1M tokens96.3%
Figures: OpenAI figures as reported by InfoQ, 10 September 2026
What you can learn from this
- A computer-use agent is a loop, not a single answer. The model receives an observation of the screen, typically a screenshot or an accessibility tree, decides on one action such as a click or keystroke, executes it, and looks again. OSWorld measures how often that loop completes real tasks in a real desktop. It is hard because small errors compound: a misread button on step three sends every later step down the wrong path, and user interfaces change in ways training data did not anticipate. Benchmarks like OSWorld therefore reward recovery from mistakes as much as first-try accuracy.
- Context windows and memory are different things. A million-token window means the model can attend to that much text at once, but attention cost grows with length and everything vanishes when the session ends. The "notes across windows" mechanism described for Codex is a form of external memory: the model writes down what it has learned and reads it back later. Agents need this because long tasks outlast any single context, and a written summary is cheaper and more reliable than re-reading everything.
- Hallucination rate matters more once a model can act. A hallucination is a confident, fluent answer that is wrong. In a chat, a human can catch it; in an agent, the wrong answer becomes a wrong click or a wrong database update. Labs reduce the rate by training models to say they do not know, by grounding answers in retrieved documents and by verifying with tools. A rate reported at 4.2 percent still means roughly one wrong answer in twenty-five, which is why verification steps remain part of any agent design.
- Preparedness frameworks tie capability to release conditions. Frontier labs publish frameworks that define capability levels in areas like cybersecurity and biology and specify safeguards that must be in place before a model at a given level ships. Classifying a model as able to find and exploit unknown vulnerabilities triggers extra controls, which is consistent with a rollout that starts with selected organisations. The idea is that the decision to release becomes a checklist rather than a judgement made under commercial pressure.
- Distribution through multiple clouds is a product feature. Offering the same model on Azure and Bedrock as well as OpenAI's own API lets enterprises keep their existing identity, billing, logging and data-residency arrangements. For a developer the practical consequence is that the same model can behave slightly differently across providers in rate limits, version pinning and available features, so the choice of endpoint is part of the architecture. Pinning a specific model version at each endpoint is the simplest defence against silent behaviour changes.
We teach this
How to use this in practice
- Build a five-step computer-use loop of your own. Install Playwright with
pip install playwrightandplaywright install chromium, then write a script that opens a public test page, takes a screenshot, sends the image and a one-line task to an LLM API that accepts images, parses a click coordinate or a text-entry instruction from the reply, executes it, and repeats up to five times. Log every observation and action to a file. Done is a run that types a term into a search box and submits it without you touching the keyboard, and a log you can read back to see where the model hesitated. - Put a number on long-context cost. Take a long document you have, such as a 150-page PDF, extract the text and count tokens with a tokenizer library such as
tiktoken. Multiply by your provider's per-token input price to get the cost of one full-context query, then estimate the cost of chunking the document and retrieving only the top five passages for the same question. Done is two figures side by side and a one-line rule for when each approach makes sense. - Run a small hallucination check. Write twenty questions where you know the answer, including five that have no answer at all or refer to things that do not exist. Ask a model each one and score the replies as correct, wrong or abstained. Repeat with a system prompt that tells the model to say when it is unsure. Done is a table with those three counts for both runs and a note on which of the unanswerable questions still drew a confident reply.
- Draw the agent loop with its failure points. Sketch observe, plan, act and verify as a cycle, show where a notes file or memory store sits, and mark the step at which a misread screen would derail the run. Then add one check at that step, such as confirming the page title before clicking. Done is a diagram you could hand to someone building their first agent.
Sources
Our reporting is an original summary; full coverage is at the links above.
Don't just read about it — build it.
Square 1 teaches the skills behind the headlines, with every line of your work graded by AI. Find your starting point in 3 minutes.
Get your free skill report