GPT-6 Astra arrives: price, access, benchmarks, and the cybersecurity limits
OpenAI has released GPT-6 Astra, a frontier model aimed at complex work that combines reasoning with tools. The company is positioning it for software engineering, browser and computer operation, research, and other long-running professional tasks. Astra is available first to a limited group, with broader product and API access scheduled to follow over the coming days.[1]
What OpenAI is launching
Astra is not presented as a narrow coding model. OpenAI describes it as its new general-purpose flagship, built to work across code, documents, browsers, visual interfaces, and scientific problems. Its context window reaches 1.05 million tokens, while a single response can contain up to 128,000 output tokens. The model supports structured outputs, function calling, web and file search, code interpreter, computer use, MCP, image generation, and hosted shell tools.[2]
For developers, the important shift is the emphasis on completing multi-step work rather than producing an isolated answer. OpenAI says Astra is better at operating software, maintaining context through longer tasks, and adjusting when requirements change. Those claims make the model particularly relevant to coding agents and automation systems where planning, tool use, and recovery from errors matter as much as code generation.[1]
Price and availability
OpenAI lists Astra API pricing at $10 per million input tokens and $50 per million output tokens. The model does not currently support fine-tuning. Its published knowledge cutoff is April 30, 2026.[2]
| Item | Current information |
| --- | --- |
| API input | $10 per million tokens |
| API output | $50 per million tokens |
| Context window | 1,050,000 tokens |
| Maximum output | 128,000 tokens |
| Fine-tuning | Not supported |
The rollout begins with organizations in OpenAI’s Trusted Access Program. OpenAI says access through the API and ChatGPT Plus, Pro, Business, and Enterprise plans will expand over the following days. The company also names Microsoft Azure and AWS Bedrock as distribution channels.[1][2]
The benchmark claims
OpenAI reports 98% on FrontierMath Tier 4, 99.9% on ARC-AGI-3, and 100% on ExploitBench. These are vendor-published results, so they should be treated as launch claims until independent evaluations reproduce them under comparable conditions.[1]
The numbers nevertheless show where OpenAI expects Astra to differentiate itself: advanced mathematics, general problem-solving, software interaction, and cybersecurity. OpenAI also says the model completes computer-based work more quickly than GPT-5.6 Sol while using fewer output tokens on several internal evaluations. The practical question for teams will be whether those gains survive real repositories, permission boundaries, flaky tools, and long agent sessions.[1]
Why it matters for developers
Astra expands the ceiling for workflows in which a model does more than suggest the next line of code. Its tool support and large context window make it suitable for repository-scale investigation, coordinated edits, browser-driven testing, research, and document-heavy engineering tasks. The higher per-token price means the useful comparison will be cost per completed task, not simply cost per token.[1][2]
Teams evaluating Astra should measure completion rate, review burden, latency, and total token usage on their own work. A model that costs more per token can still be cheaper if it needs fewer retries or produces smaller outputs, but OpenAI’s launch benchmarks alone cannot establish that outcome for a specific codebase.
The cybersecurity boundary
OpenAI says Astra is its first model to reach the Critical cybersecurity capability threshold under the company’s Preparedness Framework. In controlled evaluations, the model found previously unknown vulnerabilities and combined flaws into working exploit chains. OpenAI reports that two zero-day vulnerabilities discovered during testing are being disclosed to their maintainers.[3]
Those capabilities will not be generally available at launch. Advanced cybersecurity access starts with a small group of alpha testers and is expected to expand through the Daybreak Blue program for defensive work. OpenAI says it strengthened isolation, monitoring, refusal training, and controls around unauthorized actions before release.[3]
For ordinary developers, Astra’s launch is therefore both a capability story and an access-control story. The broadly available model is designed for powerful agentic work, while its most sensitive cyber functionality remains gated. That distinction will matter when teams compare benchmark headlines with what they can actually deploy in production.[1][3]