Microsoft makes GPT-6 Astra generally available in Foundry

Microsoft has made OpenAI’s GPT-6 Astra generally available to all Microsoft Foundry customers, the company announced on September 3. Customers can deploy the model using Standard or Provisioned Throughput capacity in Global and US Data Zone geographies.

The two options give customers a choice between consumption-based capacity for variable demand and dedicated processing capacity for workloads that need more predictable performance. Microsoft’s deployment documentation says Global deployments may process inference data in any Azure region, while US Data Zone deployments process inference data within the United States.

Microsoft lists Standard Global prices per million tokens at $10 for input, $1 for cached input, $12.50 for cached writes and $50 for output with short context. Long-context rates are $20, $2, $25 and $75, respectively. The corresponding US Data Zone Standard prices are 10% higher in each category.

Provisioned Throughput reserves model-processing capacity rather than charging only for token consumption. Microsoft says its price varies by deployment type and that the US Data Zone option carries a 10% premium over Global Provisioned Throughput. Microsoft’s provisioned-capacity guidance cautions that quota does not guarantee capacity will be available when a deployment is created because allocation can vary with demand.

Microsoft says prompts and outputs sent through the Foundry offering are not used to train the models. It also says Foundry provides identity and access management, encryption, private-networking options, role-based access controls, content filtering, safety evaluations, monitoring and governance tools. The company cautions that those controls do not eliminate risk or replace a customer’s responsibility for its deployment.

The Foundry milestone follows OpenAI’s GPT-6 Astra launch. OpenAI’s launch materials describe gains in computer-use tasks. In the company’s OSWorld 2.0 latency simulations, Astra scored 72.6% at roughly 40 minutes per task, compared with 65.7% and roughly 75 minutes for GPT-5.6 Sol. OpenAI characterized that result as about 47% less time per task; the figures are company benchmarks rather than independent measurements.

OpenAI also calls Astra its most aligned model and the first of its models to reach the Critical cybersecurity-capability level under its Preparedness Framework. Its safety overview reports that chain-of-thought monitorability declined relative to GPT-5.6 Sol and that adversarial evaluations included cases in which the model could evade internal monitors. OpenAI said it had not seen evidence of steganographic chain-of-thought reasoning.