Key findings
- Azure Data Zone deployments and OpenAI's regional processing both cost 10% more than the global short-context prices for GPT-6 Astra, GPT-6.1 Sol, and GPT-6 Luna.
- On Amazon Bedrock, AWS manages access, routing, and billing. Asynchronous tool calling, mid-turn steering, hosted file search, the shell tool, and remote MCP servers are not available there.
- On Azure OpenAI, Microsoft is the data processor and OpenAI cannot see prompts. On Bedrock, AWS sets retention, and setting store to false does not by itself guarantee zero data retention.
- OpenAI's Build tier starts GPT-6 Astra at 5,000 requests per minute and Azure's Tier 1 Global Standard at 1,000, both with 1,000,000 tokens per minute.
- Of the 112 public accounts GILD had read by October 26, 2026, 9 favor buying through a cloud provider, 8 of them through Azure OpenAI, and 7 describe contracting with OpenAI directly.
Choose between OpenAI's API, Azure OpenAI, and Amazon Bedrock on data processor, features, and quotas, because OpenAI and Azure Global Standard list the same GPT-6 prices. Check Bedrock against your feature list first, since it drops the most Responses API tools.
Three places, one model family
OpenAI's pricing page says AWS and Microsoft bill for OpenAI models used on Amazon Bedrock and Microsoft Azure. On Bedrock, GPT-6 Sol and Luna run through Bedrock Runtime with US and global inference profiles and through the Mantle endpoint in us-east-1. GPT-6 Astra runs on Runtime and on Mantle in us-west-2. Azure sells the same models as Azure OpenAI in Microsoft Foundry Models, in Global, Data Zone, and regional deployments.
Feature coverage
Checked against OpenAI's Bedrock guide and Microsoft's Foundry documentation on October 12, 2026.
- Not on Bedrock: asynchronous tool calling, mid-turn steering, programmatic tool calling, multi-agent, hosted file search, the shell tool, the image generation tool, and remote MCP servers.
- Agents: the Agents API runs in OpenAI's service; Bedrock offers Bedrock Managed Agents, powered by OpenAI, with the agent runtime hosted in Bedrock.
- Bedrock: hosted web search on the Mantle endpoint only.
- Azure OpenAI: the Responses API, with multi-agent and computer use on supported models, plus provisioned throughput with spillover to standard deployments.
Quotas and reliability
OpenAI's tiers rise with credit purchases. Azure's seven tiers depend on consumption, payment history, and Enterprise Agreement status. Teams report uneven Azure capacity. One team described small quotas and a bureaucratic increase process. Grab shared a 240,000 tokens-per-minute quota for data classification in 2024. Another team saw up to three minutes to first byte. One developer reported about 87% uptime for Azure-hosted GPT-5.6 Sol on OpenRouter.
Why teams buy through the cloud anyway
Contracts and data terms drive the choice more than price. One consultant reports that virtually all clients use Bedrock or Azure with zero data retention and reject direct model-provider contracts. Feature and caching gaps still matter. A European AI agent company wants OpenAI's EU endpoint because Azure lacks priority processing and vector stores. Another team measured about 96% cache hits on Azure Chat Completions for GPT-5.6 Sol and none on Responses.
Methodology
GILD read OpenAI's pricing, Bedrock, and data controls documentation and Microsoft's Foundry pricing, quota, deployment, and privacy documentation, checked October 12, 2026. Azure prices come from the Azure Retail Prices API for East US 2. Practitioner counts come from the 112 public accounts GILD collected for the OpenAI API, published between January 31, 2025 and October 15, 2026. Among them, 16 say how the writer buys OpenAI models.
Products in this piece
Sources
- Official documentation
- 14
- Public source
- 7
- OpenAI API docs: Pricing
- OpenAI API docs: OpenAI on Amazon Bedrock
- OpenAI API docs: Bedrock Managed Agents
- OpenAI API docs: Data controls in the OpenAI platform
- OpenAI API docs: Rate limits
- OpenAI API docs: Spend limits
- OpenAI API docs: Agents API
- Microsoft Learn: Foundry Models sold by Azure
- Azure Retail Prices API: Azure OpenAI GPT6 meters, East US 2
- Microsoft Learn: Deployment types for Microsoft Foundry Models
- Microsoft Learn: Azure OpenAI quotas and limits
- Microsoft Learn: Manage traffic with spillover for provisioned deployments
- Microsoft Learn: Data, privacy, and security for Models sold by Azure
- Microsoft Learn: Create and manage budgets
- Hacker News comment on clients buying through Bedrock or Azure (Oct 9, 2026)
- Hacker News comment on Azure-hosted GPT-5.6 Sol uptime on OpenRouter (Aug 18, 2026)
- Hacker News comment on Azure OpenAI latency spikes (Apr 29, 2026)
- Hacker News comment on Azure OpenAI throughput quotas (Apr 29, 2026)
- Grab Engineering: LLM-powered data classification for data entities at scale (Jul 15, 2024)
- OpenAI Developer Forum: An Azure OpenAI customer asking for OpenAI's EU data residency endpoint (Oct 13, 2025)
- OpenAI Developer Forum: Prompt caching on Azure GPT-5.6 Sol, 96% hits on Chat Completions and none on Responses (Jul 12, 2026)