GPT-6 Astra launch: API access and capabilities

OpenAI launched GPT-6 Astra as a new frontier model for computer use, browsing, software engineering, science, professional work, and cybersecurity. This note recaps that launch, then shows the GPT-6 Astra API surface TTAPI currently documents.

OpenAI calls Astra its most intelligent and aligned model to date. The announcement highlights saturated or near-saturated results on FrontierMath Tier 4, ARC-AGI-3, and ExploitBench, plus a new high on scientific terminal workflows. Those numbers are official-reported under OpenAI's eval setups, not a TTAPI bake-off.

Astra is rolling out in phases to ChatGPT Plus, Pro, Business, and Enterprise, and as gpt-6-astra on the OpenAI API, Azure, and Bedrock. Enterprise access starts off. Pro, Business, and Enterprise plans also get an Astra Pro tier on ChatGPT. Treat that product map as OpenAI's, then use the TTAPI model page for the gateway surface.

• Computer use and browsing
• Documents, slides, spreadsheets, and hosted sites
• Software engineering and long sessions
• Science and math workflows
• Higher cyber capability with tighter launch limits

OpenAI presents Astra as its strongest computer-use model: fill forms, update a CRM, organize a calendar, research a topic and draft into email or a document, inspect scientific data, stand up a site and run frontend QA, or install and debug software from the screen.

The published efficiency story is also official-reported: higher OSWorld 2.0 scores in less simulated time than GPT-5.6 Sol, a new high on Agents' Last Exam, and a faster Codex computer-use harness. Those demos live on OpenAI's products. A gateway chat request is not automatically a computer-use agent.

Astra is trained to produce slides, documents, spreadsheets, and analyses that follow a template and a house style, and to pull in only the context that matters instead of restating the whole brief. OpenAI also ties Astra to Sites in ChatGPT, so a prompt can become a hosted site, app, or game on that product.

Judgment is part of the launch story. When a missing fact would change the outcome, Astra is supposed to ask; routine gaps can be filled with a stated assumption. In Codex it can ask asynchronously and keep moving on work that does not depend on the answer. Mid-task steering should update the plan without dropping the original goal.

OpenAI calls Astra its best software-engineering model so far, with a new high on Terminal-Bench 4.0 in the published comparison. The more durable product change is how Codex handles a full context window.

Instead of only compacting a long debug or refactor into a summary, Astra in Codex can keep searchable notes across windows and look back at earlier messages and tool output. OpenAI currently exposes that as an experimental Codex setting, with a plan to make it the default for Astra. Do not assume the same memory behavior on a one-shot chat-completions call.

The science pitch is not only graduate-level Q&A. OpenAI reports a new high on GPQA Diamond and a large jump on Terminal-Bench Science, which asks an agent to run research workflows in a terminal. Combined with computer use, Astra is described as able to open scientific software, inspect data, and help decide the next measurement.

The announcement also points to further prime-gap results. Treat those as OpenAI's research claims, then decide whether your own workload is a reasoning question, a tool-using analysis job, or something the current API surface cannot host.

OpenAI says Astra is the first model to meet the Critical cybersecurity threshold in its Preparedness Framework. Official evals without production safeguards show a large jump on exploit-development and reverse-engineering benches. That is why the launch version is narrower than the raw capability.

The shipping model is framed for defenders: secure code review and patching are in scope. More advanced cybersecurity tasks, including proof-of-concept exploits, are refused. OpenAI plans a later Daybreak path for broader defensive workflows such as vulnerability validation and detection engineering. Extra checks can pause a ChatGPT or Codex task for review, or stop an API task outright.

OpenAI also leads with alignment: Astra is supposed to stay inside the authorized scope, ask when a decision is consequential, and avoid claiming tools or powers it does not have. In one official impossible-task eval, Astra stayed in bounds in every trial where GPT-5.6 Sol often did not. Written reasoning is harder to monitor than Sol's; OpenAI flags that as an open research issue, not a solved property.

Availability is phased. ChatGPT Plus, Pro, Business, and Enterprise get the model over days, not necessarily the same hour. Developers call gpt-6-astra. OpenAI documents a Fast mode on its own API at higher speed and a higher official rate. Zero Data Retention and Private Safety Processing are official API options for eligible customers. Confirm each of those on the current page that actually hosts your traffic.

TTAPI lists GPT-6 Astra on the OpenAI LLM family as gpt-6-astra. The documented path is the synchronous OpenAI-compatible chat-completions endpoint, not Codex, Sites, or a computer-use harness. That is the honest first request: one complete, consequential task you can review.

Set TTAPI_KEY in the server environment and POST model gpt-6-astra to https://api.ttapi.io/v1/chat/completions. The example uses a product-launch review so you can judge instruction following, missing-information handling, and usable structure. Confirm optional fields and current pricing in TTAPI's own reference before you ship.

curl --request POST \
  --url 'https://api.ttapi.io/v1/chat/completions' \
  --header "Authorization: Bearer $TTAPI_KEY" \
  --header 'Content-Type: application/json' \
  --data '{
  "model": "gpt-6-astra",
  "messages": [
    {
      "role": "user",
      "content": "Review this product launch plan. Identify the three highest risks, state what information is missing, and give one acceptance check for each recommendation."
    }
  ],
  "stream": false
}'

• Use model gpt-6-astra on /v1/chat/completions.
• Keep secrets out of browser code.
• Store the prompt, response, model ID, and request identifier together.
• Confirm current fields and pricing before you ship.