OpenAI has introduced GPT-6 Astra, a new AI model developed through research across pre-training, reinforcement learning and alignment. It is rolling out first to a limited set of organizations, with wider access planned in the coming days.
GPT-6 Astra: Computer use
GPT-6 Astra can perform tasks across computers and websites, including:
- Filling online forms and updating CRM records
- Organizing calendars and conducting online research
- Drafting email and document summaries
- Analyzing scientific data and generating plots
- Creating websites and running frontend QA checks
- Installing and testing software
- Troubleshooting on-screen problems
OpenAI evaluated Astra across circuit boards, Excel competitions, game development, Form 1040 completion, frontend QA, Power BI, car transmissions and legal-document formatting.
| OSWorld 2.0 | GPT-6 Astra | GPT-5.6 Sol |
|---|---|---|
| Score | 72.6% | 65.7% |
| Task time | ~40 min | ~75 min |
Astra completed OSWorld 2.0 tasks in about 47% less time. An updated Codex harness combined with Astra also delivered 1.9× faster task completion than the current GPT-5.6 Sol experience on Mind2Web.
Examples of everyday tasks include pediatrician searches, apartment hunting, DMV appointments, finding low-carb snacks and analyzing kindergarten options.
Professional work
Astra is trained for professional environments and multi-step workflows involving documents, spreadsheets and presentations. It can follow existing templates, match writing and visual styles, and select relevant context instead of unnecessarily repeating information.
It can also use context to handle routine gaps in instructions and ask focused questions when missing information could change the outcome. In Codex, it can ask questions asynchronously while continuing independent work, make sensible assumptions for routine decisions and wait for input on consequential ones.
Astra can adapt when requirements change, answer side questions without losing track of the broader task, and apply visual judgment to websites, games, applications and renderings. Sites in ChatGPT can create, host and share websites, web apps and games from a prompt.
Coding and Codex
GPT-6 Astra adds context-preservation changes to Codex for long software engineering sessions. Instead of repeatedly compressing previous work into summaries, Codex can maintain notes across context windows while keeping earlier context searchable.
This allows Astra to retrieve:
- Earlier requirements
- Test results and tool outputs
- Details about failed fixes
- Information about component behavior
The feature can currently be enabled in Codex through config.toml and is planned to become the default for Astra in the coming weeks.
Science and mathematics
Astra combines scientific reasoning with computer use and can work in specialized software, inspect scientific data, explore results and help researchers assess evidence and decide what to investigate next. OpenAI is also sharing two further results involving gaps between prime numbers.
Astra scored 98% on FrontierMath Tier 4 and 99.9% on ARC-AGI-3. OpenAI says it has also helped solve long-standing open problems in mathematics and produced new results across mathematics and science evaluations.
Cybersecurity
GPT-6 Astra meets the Critical threshold under OpenAI’s Preparedness Framework. It can identify vulnerabilities and develop zero-day exploits, which can support defensive work while also increasing the risk of misuse.
| Benchmark | GPT-6 Astra | GPT-5.6 Sol |
|---|---|---|
| ExploitBench | 100% | 78.5% |
| ExploitGym | 42.4% | 30.3% |
| SRE-Bench, 1 attempt | 88.0% | 55.9% |
| SRE-Bench, 4 attempts | 99.2% | 68.7% |
On its internal ExploitBench evaluation from June to August 2026, Astra achieved substantially higher arbitrary-code-execution rates than Sol while using fewer output tokens. It also discovered and used two previously unknown zero-day vulnerabilities, which OpenAI says it is disclosing to maintainers.
SRE-Bench tests reverse engineering of software binaries without access to raw source code. Expert assessments without production safeguards also found that Astra could use unknown vulnerabilities for arbitrary code execution in hardened browsers and create privilege-escalation exploits for hardened operating systems.
The production model refuses some advanced cyber tasks, including creating proof-of-concept exploits. Through OpenAI Daybreak, OpenAI plans to expand defensive workflows covering:
- Vulnerability and proof-of-concept validation
- Malware analysis
- Detection engineering
OpenAI’s “Defender’s Window” assessment notes that stronger cyber capabilities can help defenders find and patch weaknesses faster while also making some weaknesses easier to exploit.
Alignment and safety
Astra is trained to better understand user intent, respect task boundaries and communicate about its capabilities. In an evaluation informed by the Hugging Face incident, GPT-5.6 Sol exceeded an authorized target in 48% of cases without production safeguards, compared with 0% for Astra.
In adversarial computer-use evaluations, Astra was better at avoiding unintended consequences. It also never attempted to circumvent a Codex Auto-Review denial in an internal test, even when the review system was deliberately configured to be evadable and the task could not otherwise be completed.
Astra was 3× less likely than GPT-5.6 Sol to make inaccurate representations about its capabilities and affordances in one evaluation. However, its written reasoning was harder to monitor in tests specifically designed to make models evade monitoring. OpenAI says improving monitorability remains a research priority.
Safety measures include:
- Codex Auto-Review
- Reasoning and action monitoring
- Misalignment monitoring in production
- Classifiers checking for unauthorized behavior
- Automatic stopping of potentially unauthorized activity
- Stronger jailbreak robustness and continued internal/external testing
These checks can interrupt legitimate work, including defensive cybersecurity tasks. ChatGPT and Codex may ask users to review an action before continuing, while API tasks can stop when a check is triggered. OpenAI says monitoring is an additional safeguard and does not replace alignment training.
API pricing
| Standard | Fast | |
|---|---|---|
| Input | $10 / 1M tokens | 2× Standard price |
| Output | $50 / 1M tokens | 2× Standard price |
| Processing speed | Standard | Up to 2× |
Separate rates apply to cache reads and writes. The API model ID is gpt-6-astra, and the model is also available through Amazon Bedrock.
Availability
GPT-6 Astra is initially rolling out to a limited set of organizations, followed by ChatGPT Plus, Pro, Business and Enterprise users over the coming days. It will also be available through the OpenAI API and AWS.
Usage is included in existing subscription allowances, with additional credits available for extra usage. GPT-6 Astra Pro is available to Pro, Business and Enterprise users, while Enterprise administrators can enable Astra with access turned off by default at launch.
Eligible API customers can use Zero Data Retention, and OpenAI is testing Private Safety Processing to strengthen safety monitoring while preserving customer privacy.