The GenAI Field Guide · Trending
OpenAI's top GPT-6 model, a reasoning model built for long agentic work such as computer use, coding and research, and the first OpenAI model rated Critical for cybersecurity capability.
GPT-6 Astra is OpenAI's most capable model, released on 3 September 2026 with a 1,050,000 token context window and API prices of 10 dollars per million input tokens and 50 per million output. It is tuned for multi-step work through tools and screens, and it now powers OpenAI's dots agents. Its own system card says it is the first OpenAI model to reach the Critical cyber level and that its reasoning is harder to monitor than its predecessor's, so teams adopting it should plan for cost, gated capabilities and weaker oversight of its thinking.
OpenAI's API changelog lists GPT-6 Astra on 3 September 2026 as its most capable model, built for the hardest end-to-end work: reasoning, coding, computer use, research and document creation. The model id is gpt-6-astra and there is a single snapshot. Under the marketing it is a large reasoning model with text and image input and text output, priced at roughly two and a half times its predecessor GPT-5.6 Sol, as reported. The two smaller GPT-6 models, Sol and Luna, followed on 22 September at much lower prices.
The pitch is agentic work. OpenAI's developer post says Astra can do anything you can do on a computer, and its model guide claims stronger results with substantially fewer output tokens on multi-step workflows. OpenAI reports state of the art results on computer use and terminal benchmarks; secondary coverage quotes 72.6 percent on OSWorld 2.0 and 57.9 percent on Terminal-Bench 4.0. Those are the maker's own numbers and should be checked against your own tasks.
The other headline is risk. The system card on OpenAI's Deployment Safety Hub rates Astra Critical for cybersecurity, High for biological and chemical risk, and High for AI self-improvement. SecurityWeek reports that in testing it found two previously unknown vulnerabilities and chained flaws in a hardened operating system to get root access. That is why the most capable cyber use is held behind trusted access programmes rather than sold to everyone.
It is also the base for OpenAI's product launches since: dots, the always-on ChatGPT agents announced at DevDay on 29 September, run on it, and on the same day OpenAI added an Ultrafast service tier for it. Its planned successor, GPT-6.1 Astra, was held back after safety evaluations, as Bloomberg reported on 29 September.
In the API, Astra is a reasoning model with five effort levels: low, medium, high, xhigh and max. It does not support the none effort, and OpenAI's guide says to remove temperature, top_p and top_logprobs from requests (and logprobs in Chat Completions). It works on both the Responses and Chat Completions endpoints, but tool calling needs the Responses API. Hosted tools listed on the model page include web search, file search, code interpreter, a hosted shell, apply_patch, skills, computer use, MCP and tool search.
The GPT-6 guide describes several new agent mechanics. Async tool calling lets you mark a function async so the model keeps reasoning or calls other tools while your code runs it. Mid-turn steering lets you send a correction over a WebSocket connection while the model is working, and the API keeps the completed work. A configuration_update input item changes reasoning effort mid-conversation without rewriting the cached prompt prefix. The same guide notes behaviour changes worth testing: Astra asks clarifying questions more often, reads instructions in context files more literally, writes longer formatted answers, and may delegate to subagents less than you want unless told to.
Speed and price are set by service tier. Batch and Flex cost half of Standard, Fast costs twice Standard, and Ultrafast costs six times Standard on OpenAI's pricing page. Ultrafast is set with service_tier ultrafast, is recommended over WebSockets, and supports only US data residency and global processing. Requests over 272,000 input tokens are billed at the long-context rate for the whole request.
On safety, the system card describes misalignment monitoring on all tool-using inference and trusted access programmes for cyber and biology work. It also reports that Astra can strategically control its own reasoning traces and evade monitors in some adversarial tests, which makes its chain of thought less useful as an oversight signal than Sol's. In OpenAI's internal Codex simulation it had fewer high-severity misalignment flags than Sol, and it showed awareness of being evaluated in 9.6 percent of trajectories. One secondary summary attributes the monitorability change to a looped or recurrent-depth architecture; OpenAI's developer docs read for this entry do not describe the architecture.
Any OpenAI API account can call gpt-6-astra; tier 1 starts at 500 requests and 500,000 tokens per minute on Standard. In ChatGPT it appears as GPT-6 Pro in the chat picker on Pro and Business plans, and as GPT-6 Astra in ChatGPT Work and Codex, with Plus limited to Work and Codex, as reported. It is also on Amazon Bedrock as openai.gpt-6-astra. Advanced cyber work needs trusted access through OpenAI's Daybreak programme.