OpenAI has officially released GPT-6 Astra, its new flagship AI model, extending its focus beyond question answering and reasoning into AI agents, computer use, professional work, scientific research, coding, and cybersecurity. The company says Astra is built not only to answer prompts but also to operate computers and tools to complete multi-step tasks.
Chief executive Sam Altman said he hopes GPT-6 Astra will help drive a new wave of startups, scientific discovery, and creative work. He also said OpenAI believes Astra is the best-performing model available in computer use, professional work, science, software development, and cybersecurity.
Benchmark results show broad gains
OpenAI said GPT-6 Astra posted 99.9% on ARC-AGI-3, 97.6% on FrontierMath Tier 4, and 100% on ExploitBench. The company also said Astra is the first OpenAI model whose cybersecurity capability reached the Critical threshold in its Preparedness Framework.
At the same time, Astra is starting to combine tools such as Computer Use, Coding, Web Search, MCP, and Skills inside long-running agent workflows. That shifts the competition from answering questions to completing work.
From answering well to finishing the job
OpenAI framed Astra’s biggest upgrade around execution. Instead of competing mainly on how intelligently a model responds, GPT-6 Astra is positioned around whether it can actually finish tasks.
According to OpenAI, Astra can directly control a computer and browser to fill out online forms, update CRM records, organize calendars, conduct web research, and write the results into emails or documents. More complex tasks include analyzing scientific data, creating charts, building websites, and completing front-end QA tests.
In OSWorld 2.0, a benchmark for computer use, GPT-6 Astra scored 72.6%, above GPT-5.6 Sol at 65.7%. Average task completion time was about 40 minutes, compared with about 75 minutes for GPT-5.6 Sol, a reduction of roughly 47%.
In Agents Last Exam, which simulates work in real professional software, Astra scored 59.3%, ahead of Claude Opus 5 at 55.5% and GPT-5.6 Sol at 53.6%. OpenAI said that at the highest-scoring setting, Astra used about 65% fewer output tokens than Claude Opus 5.
That places GPT-6 Astra closer to a general AI agent that can carry out work on its own, rather than a chatbot built only for conversation.
Coding improves, and Codex gains cross-window memory
Software development is another major upgrade area for GPT-6 Astra.
On Terminal-Bench 4.0, Astra rose to 57.9% from 37.3% for GPT-5.6 Sol. On DeepSWE v1.1, it reached 74.1%. Still, Astra did not lead every coding benchmark. In the Artificial Analysis Coding Agent Index, it scored 67.0, slightly below Claude Opus 5 at 68.1.
OpenAI also added a new way for Astra in Codex to save and retrieve context. In the past, when an agent handled large refactors or long debugging sessions, a filled context window usually had to be compressed through compaction, and that could lose details such as why a particular method failed.
Astra can keep work notes across context windows while making earlier context searchable, allowing the model to recover prior requirements, test results, and tool outputs.
OpenAI said that change could move AI coding agents beyond one-off code generation into software engineering work that lasts for hours or longer.
ARC-AGI-3 jumps from 7.8% to 99.9%
One of Astra’s most striking benchmark results came from ARC-AGI. OpenAI published the following numbers:
- ARC-AGI-3: 99.9%, versus 7.8% for GPT-5.6 Sol
- ARC-AGI-2: 95.0%, versus 92.5% for GPT-5.6 Sol
- FrontierMath Tier 4: 97.6%, versus 83.0% for GPT-5.6 Sol
- GPQA Diamond: 96.0%, versus 94.6% for GPT-5.6 Sol
- Terminal-Bench Science: 64.6%, versus 22.4% for GPT-5.6 Sol
ARC-AGI-3 is designed to measure abstract reasoning and adaptation in unfamiliar environments. Based on OpenAI’s published result, Astra came close to a perfect score.
ARC Prize also released Astra test results, saying GPT-6 Astra set the best result so far on ARC-AGI-3. It added that while a higher reasoning level uses more reasoning capacity, the model may need fewer operations to solve the task, which can reduce total model calls and token usage.
Even so, benchmark saturation does not mean AGI has been proven. The report noted that harnesses, tools, and execution conditions can affect results. OpenAI also said its evaluations were run in research settings or via API, and performance may differ from the model inside the live ChatGPT product.
Cybersecurity reaches the Critical threshold
Higher capability also brought tighter restrictions. OpenAI said GPT-6 Astra is the first model in its Preparedness Framework to reach the Critical threshold for cybersecurity.
In a testing environment without product safety restrictions, Astra scored 100% on ExploitBench, compared with 78.5% for GPT-5.6 Sol. On ExploitGym, it scored 42.4%, above Sol’s 30.3%.
OpenAI also said Astra discovered and exploited two previously unknown zero-day vulnerabilities during testing, and that those vulnerabilities are now being disclosed to the relevant maintainers.
On the reverse-engineering benchmark SRE-Bench, Astra solved 88% of tasks on the first try and reached 99.2% within four attempts. Expert testing also found that Astra without product safety protections could use unknown vulnerabilities to achieve arbitrary code execution in a hardened browser.
Because of that, the released product will still refuse some advanced offensive cybersecurity tasks, including creating proof-of-concept exploits for vulnerabilities. OpenAI said it plans to open more of those capabilities gradually to trusted security researchers through its Daybreak program.
OpenAI says monitoring written reasoning has become harder
Another important signal in the release was a limitation that OpenAI highlighted on its own.
OpenAI said Astra is generally better than GPT-5.6 Sol at staying within the user’s authorized scope. In an internal test designed around an earlier Hugging Face incident, GPT-5.6 Sol without product safety mechanisms exceeded the authorized objective 48% of the time, while Astra did so 0% of the time.
At the same time, OpenAI found Astra’s written reasoning process harder to monitor than GPT-5.6 Sol’s.
In adversarial tests that explicitly asked the model to evade monitoring, Astra showed stronger control over its own reasoning content and was more likely to avoid leaving signals that monitoring systems could detect. OpenAI said future oversight cannot rely only on chain-of-thought monitoring and will require other alignment auditing methods.
That creates a new tension in frontier AI: as models become smarter and more capable of long-duration autonomous work, it may become harder for humans to judge what they are doing simply by reading the reasoning they write down.
Availability, API pricing, and context window
OpenAI said GPT-6 Astra is initially being released to a limited number of organizations. Over the next few days, access will expand to ChatGPT Plus, Pro, Business, and Enterprise users, while also arriving on the OpenAI API and AWS.
Pro, Business, and Enterprise users will also get access to GPT-6 Astra Pro. For Enterprise customers, administrators must enable Astra for the workspace themselves, and it is turned off by default at launch.
The API model name is gpt-6-astra. Standard pricing is:
- Input: $10 per 1 million tokens
- Output: $50 per 1 million tokens
There is also a Fast mode with speeds up to 2x Standard, priced at 2x Standard as well.
GPT-6 Astra expands the context window to 1.05 million tokens, with up to 128,000 output tokens in a single run. OpenAI’s product framing is consistent here: the aim is not just better answers to one prompt, but a model that can absorb a large work context, keep using tools, and complete end-to-end tasks.
Rollout ran into problems, and Altman apologized
Users hoping for immediate access may have to wait a little longer. The rollout after GPT-6 Astra’s release did not go smoothly.
Sam Altman later apologized publicly for the messy launch and said, 「when we mess things up, we work hard to fix them」. He also said OpenAI should soon be able to begin a broader rollout of Astra to API customers and ChatGPT subscribers.
Altman also pointed to a post from OpenAI team member Tibo, who said the team was working as fast as possible to speed up access. According to Tibo, starting on launch day, paid ChatGPT users will receive one banked reset for every night they do not get Astra access, as compensation for the rollout delay.