OpenAI has begun rolling out GPT-6 Astra to business customers, framing its newest frontier model less as a chatbot and more as a worker that operates software directly. In a company post on the next generation in intelligence for work, OpenAI pitched Astra at organisations that want multi-step tasks finished rather than answers drafted.
Key takeaways
GPT-6 Astra is priced at $10 per million input tokens and $50 per million output tokens, rising to $20 and $75 for long-context requests.
Astra scored 62.7% on the ARC-AGI-3 reasoning benchmark with a standard harness and 99.9% with an enhanced one, against 7.8% for GPT-5.6 Sol and 30% for Claude Opus 5.
Rollout starts with OpenAI's Daybreak enterprise programme before reaching ChatGPT Plus, Pro, Business and Enterprise tiers, the API, Microsoft Foundry, GitHub Copilot and AWS Bedrock.
What computer use actually means
The feature OpenAI leads with is computer use. Instead of relying on purpose-built API integrations for every system a company runs, Astra is meant to drive ordinary software the way a person does, clicking through browsers, filling forms, editing spreadsheets and assembling finished documents and presentations.
OpenAI president Greg Brockman singled the capability out as the most significant change in the release, while researcher Mia Glaese argued it has moved from an aspirational training target to something producing real value. That framing matters commercially: an AI agent able to operate legacy internal tools removes the integration work that has stalled most enterprise agent pilots.
Astra also ships with a one-million-token context window, aimed at long-horizon jobs that span many documents or an entire terminal session.
How Astra prices against the field
Astra is expensive by current standards. At $10 per million input tokens and $50 per million output tokens, it sits at the top of the market at a moment when Chinese labs are cutting list prices aggressively. Cached input drops to $1.00 and cache writes cost $12.50, so heavy reuse of the same context softens the bill considerably.
Subscription customers are handled differently. Astra usage is folded into existing ChatGPT allowances, with extra credits sold on top. Business and Enterprise workspaces do not receive the model automatically; an administrator has to switch it on. Beyond OpenAI's own API, the model is generally available in Microsoft Foundry, in GitHub Copilot for paid tiers, and announced for AWS Bedrock.
The harness caveat behind the headline score
That 99.9% ARC-AGI-3 figure comes with a footnote. The result used an enhanced harness, while the standard harness produced 62.7%. Both numbers dwarf GPT-5.6 Sol's 7.8%, but only one reflects the setup most developers will actually run.
Astra reasons using a technique OpenAI describes as recurrent depth, sometimes called looped transformers. It is more compute-efficient, but it obscures part of the chain of thought, making the model harder to audit. For regulated industries, that is an awkward property to buy into.
Why the cybersecurity restrictions are unusual
OpenAI says Astra crosses its critical cybersecurity capability threshold, meaning it can locate and exploit unknown flaws without human direction. It scored 100% on ExploitBench against 78.5% for its predecessor.
The company delayed the release to add safeguards after a July incident involving Hugging Face infrastructure, restricted Daybreak customers to defensive security work, and submitted the model to the US government for review under a voluntary framework. General users receive a version that refuses advanced offensive tasks. The training run itself spanned more than 100,000 GPUs at the Stargate site in Texas.
What comes next
The commercial question is whether computer use survives contact with messy corporate software. OpenAI has already leaned on Astra internally for research, crediting an earlier build with ten new mathematical results verified in Lean. Reproducing that reliability inside an accounts-payable workflow is a different problem, and the answer will arrive in deployment logs rather than benchmark tables.
FAQ
How much does GPT-6 Astra cost?
API access is $10 per million input tokens and $50 per million output tokens at standard context length. Long-context requests are billed at $20 and $75, while cached input falls to $1.00 per million and cache writes cost $12.50.
Can anyone use GPT-6 Astra today?
Not immediately. OpenAI started with enterprise customers in its Daybreak programme, then widened access to ChatGPT Plus, Pro, Business and Enterprise subscribers, the API and AWS. Business and Enterprise workspace administrators must enable the model manually.
Why can't developers see Astra's full reasoning?
Astra uses a recurrent-depth architecture that loops through transformer layers instead of emitting every intermediate step as text. The approach saves compute but hides part of the chain of thought, so some of the model's internal reasoning is not inspectable.
