OpenAI released GPT-6 Astra on September 3, 2026, as the top tier of its GPT-6 family. It is available in ChatGPT for paid plans, in the OpenAI API as gpt-6-astra, and through Microsoft Azure and Amazon Bedrock, at $10 per million input tokens and $50 per million output tokens. It is also the first OpenAI model rated Critical for cybersecurity capability under the company's Preparedness Framework, which shapes how it behaves and who can use it for what.
This overview covers what Astra is, how it was deployed and where it fits now that cheaper GPT-6 tiers exist. Two earlier articles cover its pricing in depth: how the 272,000-token threshold reprices requests and how Astra's pricing structure compares with Claude Fable 5.1.
OpenAI's announcement describes Astra as its most capable model, built from work across pre-training, reinforcement learning and alignment, and claims state-of-the-art results on computer use, browsing, software engineering, cybersecurity, science and professional work. The model page lists:
Just under three weeks later, OpenAI added cheaper tiers below it: GPT-6 Sol and GPT-6 Luna on September 22, and GPT-6.1 Sol on September 29. That changes the practical question from "should a team use Astra?" to "which tasks justify paying five times the GPT-6.1 Sol rate?"
According to the announcement, Astra rolled out first to a limited set of organizations and then, over the following days, to ChatGPT Plus, Pro, Business and Enterprise users and to the API, Azure and Bedrock. Some deployment details matter for administrators:
| Per million tokens (USD) | GPT-6 Astra | GPT-6.1 Sol |
|---|---|---|
| Input | $10.00 | $2.00 |
| Cached input | $1.00 | $0.10 |
| Cache writes | $12.50 | $2.50 |
| Output | $50.00 | $10.00 |
Source: OpenAI model pages for GPT-6 Astra and GPT-6.1 Sol, checked October 10, 2026.
Prompts above 272,000 input tokens are billed at 2x input and cache rates and 1.5x output for the whole request. The threshold article works through the budgeting consequences.
OpenAI's safety overview says Astra is its first model to reach the Critical level of cybersecurity capability, meaning that with the right tools and access it can find previously unknown security flaws and develop ways to exploit them across well-protected systems without step-by-step human guidance. OpenAI says it responded with stronger protections against harmful cyber actions, more robust jailbreak resistance, and, for users flagged as potentially high risk, a more conservative refusal boundary covering a broader range of dual-use requests.
For legitimate security and infrastructure teams, the operational consequence is that some requests may be refused or handled more cautiously than on earlier models, and that behavior can depend on account-level risk signals. Teams whose work involves vulnerability research should test their real workflows early, and review OpenAI's GPT-6 Astra system card for the specific safeguards.
OpenAI's headline claims include a 98% score on FrontierMath Tier 4, 99.9% on ARC-AGI-3 and 100% on ExploitBench, all of which it describes as saturating those benchmarks. For computer use, OpenAI's latency simulations on OSWorld 2.0 report a partial score of 72.6% on the offline set at roughly 40 minutes per task for Astra, against 65.7% at roughly 75 minutes for GPT-5.6 Sol. On alignment, OpenAI says that on an evaluation of whether a model exceeds its authorized scope on a difficult task, GPT-5.6 Sol without production safeguards did so 48% of the time and Astra 0%.
These are OpenAI's results, and this article does not reproduce them. Saturated benchmarks also say less about differences between top models than unsaturated ones do. Later releases from OpenAI and Anthropic compare themselves with Astra, generally claiming similar results at lower cost per task: OpenAI for GPT-6.1 Sol, and Anthropic for Claude Opus 5.5 in its Opus 5.5 announcement (see the Claude Opus 5.5 guide). OpenAI's own guidance after GPT-6.1 Sol is to keep Astra for the hardest tasks.
Teams with hard, high-value tasks where a failed attempt costs more than the tokens, such as complex scientific analysis or long computer-use workflows, are the clearest fit. OpenAI's own GPT-6.1 Sol announcement points to Astra for the most difficult research tasks.
Teams already on Astra for everything should run their task set against GPT-6.1 Sol and route only the tasks that fail there to Astra.
Teams comparing frontier models can read the Claude Fable 5.1 vs GPT-6 Astra comparison, which covers how the two price long contexts and caching, and the full list of September–October releases.
When was GPT-6 Astra released?
September 3, 2026, with a staged rollout over the following days.
How much does GPT-6 Astra cost?
$10 input, $1 cached input, $12.50 cache writes and $50 output per million tokens at standard rates, per OpenAI's model page (checked October 10, 2026). Requests above 272,000 input tokens cost more.
What does "Critical" cybersecurity capability mean?
It is the highest level in OpenAI's Preparedness Framework for cyber capability. OpenAI says it means the model can find and exploit previously unknown flaws in well-protected systems without step-by-step guidance, and that it added stronger safeguards in response.
Discover more content:
