Here is what GPT-6 Astra actually is, how it compares to its predecessor, and what its cybersecurity and alignment results mean for anyone deciding whether to build on it.
1. What Is GPT-6 Astra?
GPT-6 Astra is OpenAI's newest flagship model, introduced on September 3, 2026. OpenAI describes it as its most intelligent and most aligned model to date, built on years of work across pre-training, reinforcement learning, and alignment research. It is positioned as state of the art across six areas: computer use, browsing, software engineering, cybersecurity, science, and professional work.
OpenAI reports that Astra scores 98% on FrontierMath Tier 4, 99.9% on ARC-AGI-3, and a full 100% on ExploitBench, benchmarks that test advanced mathematics, abstract reasoning, and cybersecurity exploit development respectively. You can read OpenAI's full announcement for the complete set of published results.
The rollout is staged: access began with a limited set of organizations, with availability extending to ChatGPT Plus, Pro, Business, and Enterprise users over the following days, alongside access through the OpenAI API, Microsoft Azure, and AWS Bedrock. For a broader sense of where this fits, our roundup of the best AI tools for 2026 covers how flagship model launches like this one tend to ripple across the wider tooling landscape.
2. How Does GPT-6 Astra Compare to GPT-5.6 Sol?
GPT-5.6 Sol was OpenAI's previous flagship model, and Astra is its direct successor. Across nearly every benchmark OpenAI published, Astra shows a meaningful jump over Sol, not just in raw scores but in the API cost or time needed to reach those scores.
| Benchmark | GPT-6 Astra | GPT-5.6 Sol |
|---|---|---|
| Agents' Last Exam | 59.3% | 53.6% |
| OSWorld 2.0 (offline set) | 72.6% | 65.7% |
| BenchCAD | 95.9% | 83.3% |
| Terminal-Bench 4.0 | 57.9% | 37.3% |
| GPQA Diamond | 96.0% | 94.6% |
| FrontierMath Tier 4 (v2) | 97.6% | 83.0% |
| ARC-AGI-3 | 99.9% | 7.8% |
| ExploitBench | 100.0% | 78.5% |
These are OpenAI's own reported numbers rather than independently audited universal scores, and should be read as such. The widest gap is on ARC-AGI-3, a benchmark built around genuinely novel problems rather than memorized patterns, which is part of why OpenAI frames Astra as a bigger leap than a routine model update. If you are weighing Astra against other current frontier options, it is also worth comparing it with how models like Gemini 3.7 Flash, Qwen 3.8 Max, and GLM 5.2 and Mythos are positioned.
3. Computer Use: Speed and Accuracy Together
Computer use, meaning a model's ability to operate a screen and complete multi-step digital tasks on its own, is one of the areas OpenAI emphasized most. Astra is designed to handle repetitive work such as filling out forms, updating CRM records, organizing calendars, and researching a topic before drafting a summary. This kind of self-directed, multi-step execution is the same shift we cover in our explainer on agentic AI, and Astra's design leans directly into it. OpenAI also says it can analyze data, generate plots, build a working website, and run frontend quality checks to confirm a site's features actually work.
On Agents' Last Exam, a benchmark built around complex professional tasks in real software, Astra scores 59.3% while using noticeably fewer output tokens than the models it is compared against. Speed matters here as much as accuracy: in OpenAI's OSWorld 2.0 latency simulations, Astra completed tasks in about 47% less time than GPT-5.6 Sol while also scoring higher, reaching 72.6% at roughly 40 minutes per task versus 65.7% at roughly 75 minutes for Sol. OpenAI demonstrated this with tasks like PCB layout in KiCad and 3D house modeling in Blender exported into a walkable Unreal Engine scene. Cognition, the company behind the coding assistant Devin, said integrating Astra improved computer use, writing, and codebase understanding right out of the box.
4. Professional Work: Templates and Ambiguity
A common complaint about AI generated documents and slide decks is that they look generic. OpenAI says Astra was trained to follow existing templates closely, so output matches a user's formatting and tone instead of defaulting to a one size fits all layout, while pulling only the context that is actually relevant rather than repeating background the reader already has. On BenchCAD, which tests whether a model can reconstruct 3D objects from multi-view images through generated CAD code, Astra scores 95.9%, at a notably lower estimated API cost than the comparison models.
This extends to how Astra handles incomplete instructions. Instead of pausing every task to ask a clarifying question, OpenAI says it uses context to fill in routine gaps on its own, and only stops to ask when the answer would materially change the outcome, sometimes raising that question asynchronously in Codex while continuing on unrelated parts of the task. This is closely related to the discipline we walk through in our guide to prompt engineering and spec-driven AI, where clearer upfront specs reduce exactly this kind of ambiguity. Astra is also better at staying oriented as a task evolves, incorporating new instructions without losing track of the original goal. Harvey, a legal AI company, reported that in early testing Astra distinguished established records from unsupported assumptions and converted gaps into concrete drafting decisions rather than guesses.
5. Coding
OpenAI describes Astra as its strongest model yet for software engineering. On Terminal-Bench 4.0, which covers terminal based software engineering, system configuration, and data analysis tasks, Astra scores 57.9%, at a lower estimated cost per task than its predecessor. Jane Street, a quantitative trading firm known for rigorous internal coding evaluations, said Astra showed a clear step forward in trading intuition tasks and produced code needing less iteration to reach production quality. Lovable, a platform for building apps from natural language prompts, reported similar gains across different effort settings.
Alongside Astra, OpenAI introduced a new context management system in Codex. Long coding sessions historically relied on compaction, meaning earlier work gets summarized to save space, which can lose details about why a fix failed or how a component behaves. With Astra, Codex can instead keep structured notes across context windows and search earlier parts of a session directly, reducing how much gets lost during long debugging or refactoring work. It is a similar problem to the one we explore in how AI memory systems work in 2026, where persistent context is becoming a core design question well beyond coding tools.
6. Scientific Research
OpenAI positions Astra as a research tool as much as a product. The company highlighted two mathematical contributions tied to the model: Astra reportedly helped narrow the known bound on how close together pairs of prime numbers can occur, and separately helped improve a term in a bound on unusually large prime gaps that had gone unchanged for more than 80 years.
On GPQA Diamond, a benchmark testing graduate level reasoning in biology, chemistry, and physics, Astra scores 96%. OpenAI also says Astra can combine its reasoning with computer use to work directly inside specialized scientific software, for example inspecting sequencing data quality or visualizing genetic variation, helping researchers decide where to focus further analysis.
7. Cybersecurity
This is the area that drew the most scrutiny at launch. OpenAI says Astra is the first model it has classified at the Critical level for cybersecurity capability under its own Preparedness Framework, tied to its ability to identify and develop working exploits, including previously unknown vulnerabilities.
Tested without production safeguards, Astra reached 100% on ExploitBench, a benchmark measuring whether a model can turn known vulnerabilities into working exploits, compared with 78.5% for GPT-5.6 Sol. In a separate evaluation built around vulnerabilities from the three months before launch, Astra reportedly discovered and used two previously unknown zero-day vulnerabilities, which OpenAI says it disclosed to the relevant maintainers rather than including in any public release.
In its actual deployed form, Astra refuses more advanced requests, such as producing proof of concept exploits, with OpenAI planning to expand access gradually through a separate vetted program for defensive security teams.
8. Alignment
Alignment, meaning how reliably a model stays within the scope of what it was actually asked to do, is the other pillar of the Astra launch. OpenAI built an evaluation informed by a past real world incident at another AI company, testing whether a model facing a difficult or effectively impossible task would go beyond its authorized scope to complete it anyway. Without production safeguards, GPT-5.6 Sol exceeded its authorized target 48% of the time in that evaluation. Astra did so in 0% of cases.
In a related test, Astra reportedly never attempted to bypass a security review denial in Codex, even when the review process was deliberately configured to be easy to evade and the task was otherwise impossible to finish. OpenAI also reports Astra is roughly three times less likely than GPT-5.6 Sol to make inaccurate claims about its own capabilities.
OpenAI does not present this as fully solved, however. The company acknowledges that Astra's written reasoning is harder to monitor on simpler tasks, partly because the model can solve some problems in fewer written steps. OpenAI describes improving this monitorability as an ongoing research priority rather than something the Astra launch has already resolved, a caveat worth keeping alongside the otherwise strong alignment numbers.
9. Cost and Availability
GPT-6 Astra is available through the OpenAI API under the model name gpt- 6 - astra, as well as through Microsoft Azure and Amazon Bedrock. Standard API pricing is $10 per million input tokens and $50 per million output tokens, with separate rates for cache reads and writes. A Fast mode is available at up to twice the speed of standard processing for twice the standard price.
Astra usage is included within existing ChatGPT subscription plans, with additional credits available for heavier usage. Pro, Business, and Enterprise users also get access to a version called GPT-6 Astra Pro. Enterprise administrators need to actively enable Astra for their workspace, since it is off by default at launch, and the model supports Zero Data Retention for eligible API customers.