GPT-6 Astra reaches broad API rollout with 1.05M context — and enterprise computer-use controls

OpenAI's September 15 work-focused Astra release adds a control and evidence layer on top of the broad API rollout already captured by BTN. Enterprise administrators can restrict Astra to approved websites and desktop applications, manage uploads and downloads, and control browsing history; ChatGPT Work and Codex can also require confirmation before consequential actions and automatically review potentially unsafe tool calls. OpenAI now reports 57.9% on Terminal-Bench 4.0 versus 37.3% for GPT-5.6 Sol and 55.8% for Claude Fable 5.1, with estimated API cost per task about 9% lower than Sol and 63% lower than Fable 5.1 under OpenAI's test assumptions. OpenAI also says Astra produced unintended outcomes 89% less often than Sol and 74.7% less often than Fable 5.1 on an internal computer-use safety benchmark, and identifies Astra as its first model to reach the Critical cybersecurity capability threshold under the Preparedness Framework. The underlying API product remains `gpt-6-astra` with a 1,050,000-token context window, 128,000 output tokens, short-context Standard pricing of $10/M input and $50/M output, higher rates above 272K input tokens, and Fast/Batch/Flex tiering.

Astra now combines three operational questions that teams previously had to evaluate separately: raw capability, cost per completed task and how much computer access the model should receive. A model that can use websites and desktop software more autonomously is more useful for long workflows, but also needs explicit scope, confirmation and audit boundaries. The new admin controls therefore matter as much as another benchmark gain. For API builders, long-context pricing and the API's stop-on-safety-interruption behavior still require resumable workflows; for enterprise users, website/app allowlists and file-transfer controls create a practical way to roll computer use out gradually rather than choosing between full access and no access.

OpenAI documents `gpt-6-astra` with a 1.05M-token context window and 128K maximum output. Standard short-context pricing is $10/M input and $50/M output, with separate cached-input and cache-write prices. Input above 272K moves the request onto higher long-context rates, so large prompts need explicit cost modeling rather than assuming one flat price.

OpenAI's September 15 material reports 57.9% on Terminal-Bench 4.0 versus 37.3% for GPT-5.6 Sol and 55.8% for Claude Fable 5.1. Under OpenAI's estimated API-cost assumptions, Astra is about 9% cheaper per task than Sol and 63% cheaper than Fable 5.1 on that benchmark. Customer evaluations hosted by OpenAI also report fewer steps, fewer inference calls or higher long-horizon pass rates, but these remain vendor/customer-produced results rather than independent production measurements.

New enterprise controls can restrict Astra to approved websites and desktop applications, manage uploads and downloads, and control browsing history. ChatGPT Work and Codex add confirmation policies and automated review of potentially unsafe or unauthorized tool calls. This lets organizations begin with narrow computer access and expand it over time instead of treating computer use as an all-or-nothing permission.

OpenAI says Astra is the first model to reach the Critical cybersecurity capability threshold under its Preparedness Framework. The work release also reports that Astra produced unintended outcomes 89% less often than Sol and 74.7% less often than Claude Fable 5.1 on an internal computer-use safety benchmark. Those are first-party measurements, so they are evidence of OpenAI's deployment case rather than independent proof that every enterprise workflow is safer.

Batch and Flex cost half of Standard. Fast costs twice the applicable Standard rate and OpenAI says it can deliver up to around twice the speed. Fast is not available with EU data residency, so globally deployed applications may need a routing policy rather than one universal tier choice.

ChatGPT and Codex can pause work and surface it for review. OpenAI says an interrupted API task stops instead. Applications therefore need to persist intended actions and side-effect state so a human-approved retry can resume safely without duplicating work.

Astra in Codex can experimentally maintain notes across context windows while older windows remain searchable. This moves long-running-agent memory toward explicit retained state plus retrieval rather than relying only on the active token window. Codex can also ask asynchronous clarification questions while continuing unrelated work.