GPT-6 Astra

GPT-6 is officially GPT-6 Astra

GPT-6 Astra is OpenAI's flagship reasoning model released on September 3, 2026. Its official API model ID is gpt-6-astra. OpenAI positions it for complex reasoning, software engineering, computer use, research, and document creation, with an emphasis on carrying multistep work from an initial request to a finished result. The former GPT-6 and Astra rumor entries in this catalog now resolve to this single official product identity.

Specifications and modalities

• Context window: 1,050,000 tokens
• Maximum output: 128,000 tokens
• Knowledge cutoff: April 30, 2026
• Input: text and images
• Output: text
• Reasoning effort: low, medium, high, xhigh, and max; none is not supported

OpenAI has not disclosed total parameters, active parameters, or architecture details, so those catalog fields remain empty. The documentation confirms multilingual capability but does not provide a verified language count or complete BCP-47 list.

Agent and tool capabilities

GPT-6 Astra supports streaming, function calling, Structured Outputs, and Responses API tools including web search, file search, image generation, Code Interpreter, hosted shell, Apply Patch, Skills, computer use, MCP, and tool search. New platform capabilities include asynchronous tool calling, mid-turn steering over WebSockets, and configuration_update items that change reasoning effort while preserving the cached prompt prefix. Chat Completions is supported, but tool calling requires the Responses API.

Pricing

Standard short-context rates per million tokens are $10 input, $1 cached input, $12.50 cache writes, and $50 output. When a request contains more than 272K input tokens, the entire request uses long-context rates of $20 input, $2 cached input, $25 cache writes, and $75 output. Batch and Flex cost 50% of Standard rates, while Fast mode costs twice the applicable rate. Eligible regional processing may add 10%.

Availability

GPT-6 Astra shipped on September 3, 2026 as a staged rollout. Access went first to approved customers in Daybreak, OpenAI's application-only cybersecurity program, with ChatGPT Plus, Pro, Business, and Enterprise plans plus the OpenAI API (model ID gpt-6-astra), Amazon Bedrock, and Microsoft Azure following over the coming days. Capability differs by channel: the standard API and ChatGPT builds refuse advanced offensive tasks such as exploit discovery, while vetted defenders receive broader access through Daybreak Blue. Because the rollout is still staged, this catalog labels the model Preview; it can move to GA once broad availability is confirmed.

Safety classification

GPT-6 Astra is the first model OpenAI has designated “Critical” for cybersecurity under its Preparedness Framework, reflecting its finding that the model can autonomously discover previously unknown security weaknesses and build working exploits against well-defended systems without a human directing every step. OpenAI shipped it with additional safeguards and tiered access, and says it was pre-trained on more than 100,000 GPUs in its largest training run to date.

Migration notes

Set the API model to gpt-6-astra. Workloads that used none or minimal reasoning should start at low. The model does not accept custom temperature, top_p, or top_logprobs; Chat Completions also does not accept logprobs. Use the Responses API for tool calling. Fast mode is unavailable with EU data residency and carries no latency SLA.

Official evaluations

OpenAI's GPT-6 Astra launch page publishes 40 results across nine groups: computer use, professional work, coding, academic tasks, science and health, cybersecurity, alignment, long context, and abstract reasoning. DataLearner records every row and preserves conditions such as OSWorld 2.0 offline partial, ScreenSpot-Pro without tools, HLE with tools, length-adjusted HealthBench Professional, and the two MRCR context ranges. For ARC-AGI-3 this catalog records the ARC Prize–verified 62.7% (Semi-Private set, Standard harness, max effort, $26,098) rather than the 99.9% quoted on OpenAI's launch page. That figure comes from OpenAI's own Provider Adapter harness (high effort, $18,817), which preserves opaque reasoning state between requests and is therefore not comparable across providers. Both are state-of-the-art results; only the comparable one is used in rankings. OpenAI says each value is the best score across evaluated effort settings and that runs used either its research environment or the API, so the rows are labeled “mode unspecified” rather than uniformly described as max effort. Computer-use safety, its AutoReview variant, circumvention, ExploitGym honeypot, and hallucination are lower-is-better metrics.

Official sources

• OpenAI: GPT-6 Astra launch and full evaluation tables
• OpenAI API: GPT-6 Astra model page
• OpenAI API: GPT-6 Astra guidance and migration
• OpenAI API pricing
• OpenAI API changelog for September 3, 2026