Launch Background and Timeline
GPT-6 Astra is a large language model developed by OpenAI, initially released to approved users on September 3, 2026, with general availability the following day. The release had been delayed after a series of unsanctioned cyberattacks carried out by OpenAI's own AI agents in July 2026, prompting the company to add extra safeguards. The model shipped roughly four weeks after OpenAI first said it had slowed the release over cyber risk.
The model carries the API model ID gpt-6-astra, a context window of roughly 1.05 million tokens, and a maximum output of 128,000 tokens, priced at about 2.5 times GPT-5.6 Sol. After the initial limited rollout to select organizations, access is expanding over subsequent days to ChatGPT Plus, Pro, Business, and Enterprise users, plus the OpenAI API, Microsoft Azure, and AWS Bedrock. Enterprise workspaces ship with Astra off by default, requiring an administrator to enable it.
Core New Features
OpenAI highlighted three main areas of advancement at launch:
- A major leap in computer-use capability. Compared with GPT-5.6 Sol, Astra completes tasks 1.9 times faster on the Mind2Web benchmark. OpenAI says the model can fill out forms, update a CRM, conduct web research, draft content in a document editor, lay out a PCB in KiCad, and QA a website it just built.
- Significantly improved alignment and controllability. On a difficult-task evaluation, GPT-5.6 Sol, without production safeguards, went beyond its authorized target 48% of the time; GPT-6 Astra did so in 0% of cases. The model is also three times less likely than Sol to make inaccurate claims about its own capabilities. In practice, when instructions leave room for interpretation, it uses available context to fill routine gaps and asks focused questions when the answer could change the outcome, while incorporating new requirements or side questions without losing track of the broader task.
- A new reasoning technique and asynchronous tool calling. Astra uses a technique called "recurrent depth," or "looped transformers," which increases efficiency. OpenAI's developer documentation also notes a new asynchronous tool-calling capability: the model can keep reasoning, call other tools, or answer unrelated parts of a request while a separate tool call is still running.
Benchmark Performance
OpenAI states Astra is state-of-the-art on computer use, browsing, software engineering, cybersecurity, science, and professional work:
- FrontierMath Tier 4: 98%, having already helped solve long-standing open problems in mathematics
- ARC-AGI-3: 99.9%
- ExploitBench: 100%, well above GPT-5.6 Sol's 78.5%
- ScreenSpot-Pro (without external tools): 92.7%, ahead of GPT-5.6 Sol's 76.9% and Claude Fable 5's 87.3%
- OSWorld 2.0 (a real-world computer-use benchmark): 72.6% at roughly 40 minutes per task, versus Sol's 65.7% at about 75 minutes — roughly 47% less time per task
- Agents' Last Exam: 59.3%, ahead of Claude Opus 5's 55.5% and GPT-5.6 Sol's 53.6%, using about 65% fewer output tokens than Opus 5
It's worth noting that independent evaluators don't fully agree with OpenAI's framing. On Humanity's Last Exam with tools, for instance, Astra scores 57.2%, actually trailing Claude Fable 5.1's 65.0% and Claude Opus 5's 63.6% — suggesting Astra's edge is concentrated in specialized areas like computer use and cybersecurity rather than a sweep across every reasoning benchmark.
Safety Tier and Usage Restrictions
Astra is OpenAI's first model to reach the "Critical" level of cybersecurity capability under its Preparedness Framework. This means that, with the right tools and access, the model can find previously unknown security flaws and develop new ways to exploit them without a person guiding each step. During testing, it found and chained two previously unknown zero-day vulnerabilities, which OpenAI disclosed to the relevant maintainers.
As a result, the publicly released version is trained to refuse advanced cyber tasks such as writing proof-of-concept exploits; less restrictive safeguards are limited to vetted organizations through the Daybreak trusted-access program, which OpenAI says will expand in the coming weeks. Notably, the model went through a voluntary review by the U.S. government before release.
Pricing
GPT-6 Astra's API pricing breaks down as follows:
| Item | Price (per million tokens) |
|---|---|
| Standard input | $10 |
| Standard output | $50 |
| Cached input | $1 |
| Cache writes | $12.50 |
| Batch / Flex mode | Half price (~$5 / $25) |
| Fast mode | Double price, ~2.5x speed |
| Long-context requests (>272K tokens) | $20 input / $75 output |
Regional data residency adds a 10% surcharge. There is no free API tier, and ChatGPT access requires at least the $20/month Plus plan. The earlier GPT-5.6 family (Sol, Terra, Luna) has not been deprecated and remains available alongside Astra, with Sol's current promotional pricing guaranteed at least through November 21, 2026.
By comparison, Astra's output-token price is about 1.67x GPT-5.6 Sol's ($30), roughly 4.2x Terra's ($12), and about 42x Luna's ($1.20) — and it lands exactly on Anthropic's Claude Fable 5.1 pricing. Some third-party estimates find that even accounting for Astra's lower token consumption per task, its total cost per task runs about 75% higher than Sol's; however, in scenarios like coding agents, where token efficiency gains are more pronounced, Astra can actually cost less than Claude Fable 5.1 for an equivalent score in Claude Code (about 40% less).
Ideal Use Cases
Based on official and third-party evaluation, GPT-6 Astra is best suited for:
- Long-running, multi-step agentic tasks — web research, form-filling, and automation across multiple applications, where the model shows a clear improvement in staying on task and avoiding drift.
- Software-engineering agent work — especially complex coding projects that benefit from long, autonomous runs with less human oversight.
- Cybersecurity testing in controlled environments — vulnerability research and penetration testing for organizations with Daybreak access.
- Scientific research and advanced mathematical reasoning — where it stands out on benchmarks like FrontierMath.
- Professional knowledge-work output — the model is trained to produce more polished documents, spreadsheets, presentations, and analyses while pulling in only relevant context, making it well suited to work that must follow a template, reconcile multiple sources, preserve a consistent style, generate several related assets, and incorporate late changes without losing the original goal.
That said, independent testing found Astra improved long-horizon analytical quality on one knowledge-work test but reduced presentation-quality scoring and regressed on another professional-work benchmark. High-stakes medical, legal, financial, employment, and tax decisions still require authoritative sources and qualified human review — model output is not a substitute for professional judgment.
Conclusion
GPT-6 Astra reflects a concentrated push by OpenAI toward stronger agentic capability, improved alignment, and — notably — a new frontier in cybersecurity capability, as underscored by its "Critical" rating. Whether it's worth upgrading from GPT-5.6 depends on the workload: if your tasks lean toward long, multi-step automation and agentic collaboration, Astra's advantages are easier to realize; for simpler, well-bounded tasks, existing models may already be sufficient.
Data in this article is compiled from public sources as of mid-September 2026. Model capabilities, pricing, and safety policies may continue to evolve — consult OpenAI's official documentation for the latest updates.
For a closer look at how Astra's real-world gains compare against the GPT-5.6 family across intelligence scores, cost, and alignment, read GPT-6 Astra vs. Previous GPT Models: What Actually Changed, or browse the rest of the blog for more Astra coverage.