GPT-6 Astra

What is disclosed

OpenAI released GPT-6 Astra in the API on 3 September 2026, with its system card.¹¹1OpenAI, API changelog, entries of 3 and 29 September 2026. https://developers.openai.com/api/docs/changelog ²²2OpenAI, GPT-6 Astra System Card, 3 September 2026, sections on safety overview, model data and training, prompt injection, alignment and cybersecurity capabilities. https://deploymentsafety.openai.com/gpt-6-astra

The context window is 1,050,000 tokens. A single response can use up to 128,000 of them.³³3OpenAI, GPT-6 Astra model page, read 29 September 2026. https://developers.openai.com/api/docs/models/gpt-6-astra Astra reads text and images and writes text, and its knowledge stops at 30 April 2026.³³

The system card says little about training. Astra was "trained on diverse datasets", and it learned to reason through reinforcement learning.²² Its size and architecture are not published.

What is new

GPT-5.6 Sol, the July 2026 flagship, had the same window and a knowledge cutoff of 16 February 2026.⁵⁵5OpenAI, GPT-5.6 Sol model page, read 29 September 2026, with the release date from the API changelog entry of 9 July 2026. https://developers.openai.com/api/docs/models/gpt-5.6-sol and https://developers.openai.com/api/docs/changelog

Astra is the first OpenAI model rated Critical for cybersecurity under the company's Preparedness Framework.²² With the right tools and access, the card says, it can find previously unknown security flaws and develop ways to exploit them without a person guiding each step.²²

ExploitBench gives a model 41 known flaws in the V8 JavaScript engine and five attempts at each. Astra scored 100%, even at the lowest reasoning effort tested.²² OpenAI suspects contamination. In one task Astra failed to exploit the flaw it was given, CVE-2023-6702. It then recalled CVE-2024-0517 on its own and got arbitrary code execution through that flaw instead.²²

Against indirect prompt injection, OpenAI's internal test gives Astra a defender success rate of 99.79%.²² In a simulation over more than 54,000 internal Codex tasks, it drew about half as many high-severity misalignment flags as GPT-5.6 Sol.²²

The API now watches agent work for misalignment. A check can raise a safety alert or stop the conversation for review.¹¹ Async tool calling lets the model keep working while the application runs a tool.¹¹

Using it

The id is gpt-6-astra. Chat Completions accepts it, but tool calling needs the Responses API.¹¹ Effort runs from low to max, and none returns HTTP 400.⁴ Custom temperature and top_p values are refused, and so are log probabilities.¹¹

Standard pricing is $10 per million input tokens and $50 per million output tokens, with cached input at $1.⁶⁶6OpenAI, API pricing, Standard and Ultrafast tables, read 29 September 2026. https://developers.openai.com/api/docs/pricing. Arithmetic: Ultrafast 60 / 10 = 6 and 300 / 50 = 6; GPT-6.1 Sol against Astra 2 / 10 = 0.2 and 10 / 50 = 0.2. Past 272K input tokens, the whole request is billed at twice the input rate and 1.5 times the output rate.⁶⁶ Batch and Flex cost half of Standard.³³

Ultrafast mode arrived on 29 September. It shortens the gap between output tokens and costs six times Standard, $60 for input and $300 for output.¹¹⁶⁶

GPT-6.1 Sol charges a fifth of Astra's price for input and for output.⁶⁶ OpenAI suggests running both on the same task.⁷⁷7OpenAI, model selection guide, read 29 September 2026. https://developers.openai.com/api/docs/guides/model-selection