GPT-6 Astra reaches Microsoft Foundry: $10.00 in, $50.00 out per million tokens
GPT-6 Astra is now in Microsoft Foundry, first through its Limited Access Program, with Foundry pricing published: $10.00 input and $50.00 output per million tokens on Standard Global short context.
Microsoft has brought OpenAI's GPT-6 Astra to Microsoft Foundry and published its Foundry pricing. In its announcement on 3 September the Azure account said Astra can "act across your apps, and even use a full computer." OpenAI presented Astra the same day, claiming state of the art computer use.
The Azure blog, by Foundry Models product management vice president Steve Sweetman and Naomi Moneypenny, says delivery started through the Foundry Limited Access Program and widens to participating customers in the coming days. On 1 September the same account announced Anthropic's Claude Fable 5.1 in Foundry.
Prices, per million tokens
- Standard Global, short context: input $10.00, cached input $1.00, cache write $12.50, output $50.00
- Standard Global, long context: input $20.00, cached input $2.00, cache write $25.00, output $75.00
- Data Zone US, short context: input $11.00, cached input $1.10, cache write $13.75, output $55.00
- Data Zone US, long context: input $22.00, cached input $2.20, cache write $27.50, output $82.50
Deployment is the Standard type, Global or US Data Zone, consumption based, with no reserved capacity commitment. Microsoft says real cost varies by workload and configuration.
Microsoft names the enterprise work: reproducing a complex bug and preparing a fix for developer review, building dashboards in Power BI, producing documents from existing templates, updating customer records.
Microsoft's own caveats
Microsoft says Astra can read what is on screen and act through approved interfaces, even where a workflow has no dedicated API. It leaves the claim with its owner: "OpenAI reports state-of-the-art results on selected computer-use evaluations; performance varies by task, tools, configuration, and safeguards."
The blog says a capability this direct needs bounding, and warns that content in an application may be incomplete, misleading, or designed to influence the agent's behavior. Microsoft describes Foundry's answer as defined access, approval and monitoring: scoped credentials, approved sources, human checkpoints before consequential operations, activity logs.
Microsoft lists Entra identity management, encryption, private networking, role based access control and content filtering, and says prompts and outputs are not used to train models. It adds that these capabilities do not remove risk and do not replace each organization's responsibility to choose controls for its own scenario and regulations.
Open note: GlobalFeed AI's publishing engine runs on Anthropic's Claude model, and Anthropic is a direct competitor of OpenAI.
