Skip to main content

OpenAI API vs Azure OpenAI: Same Models, Different Contracts

DoorDash and Grab keep Azure OpenAI as a second route to the same models. Buy from OpenAI for higher starting limits, a hard spend cap, and new features first; buy through Azure when your Microsoft contract or data processor rules require it.

Reading time
6 min read
Sources
27 cited
In this piece

GILD verdict

Choose the OpenAI API for features on release day and a spend cap that stops traffic.

Teams at scale buy both and use Azure OpenAI as the second route, as DoorDash and Grab do. Choose Azure OpenAI when your Microsoft agreement decides it or Microsoft must be the data processor. Price does not decide it, because Azure Global Standard lists GPT-6 at OpenAI's own rates. Measure error rates on your own Azure deployment before launch: Requesty's gateway saw 78% of Azure requests succeed in April 2026, against 98% on OpenAI. At GILD's workload of 60,000 cached tasks a month, GPT-6.1 Sol costs US$1,632 on both, and US$1,795.20 on Azure's Data Zone.

Teams that run both use Azure OpenAI as a second route to the same models: DoorDash falls back to Azure when OpenAI's endpoint fails, and Grab routes OpenAI and Azure separately behind one AI Gateway that more than 300 use cases had joined by February 2025. Teams starting fresh get more from OpenAI directly, beginning with five times Azure's starting request rate on GPT-6 Astra. Azure OpenAI, now Azure OpenAI in Microsoft Foundry Models, sells the same GPT-6 models on Microsoft's infrastructure and contract. Global Standard deployments list at OpenAI's own prices, and Data Zone deployments cost 10% more.

What teams report

Run both 1

  1. DoorDash

    DoorDash calls OpenAI's endpoint with an Azure endpoint as its fallback, and the Claude API with Amazon Bedrock as its fallback, so product services stay up; its gateway then added workspaces that attribute usage costs to each department (InfoQ talk, Oct 3, 2026).

    “We actually had to go understand what the fallback should be, what kind of routing should be. At one point, I remember we had OpenAI endpoint, but we also had a fallback on Azure endpoint. Or, we had Claude model API endpoint, then we had a fallback on Bedrock, so that the services can rely on us from the reliability point of view. We went in that direction. We quickly realized that as the usage is going up, then we need a way to track, because now finance got involved. Like, you guys have one API key for multiple teams, how do I track who is using what? We quickly realized we need workspaces where we have cost attribution. We can actually drill down on who is using how much. We can attribute cost back to those departments. That's what basically we did with LLM Gateway.”
    Siddharth Kodwani, Software Engineer, GenAI Platform, DoorDashInfoQ, , recording

Adopted 1

  1. GrabOn Azure OpenAI

    Grab runs column-level data classification on Azure OpenAI with every deployment sharing one 240K tokens-per-minute quota; it scanned more than 20,000 data entities in its first month and estimates about 360 man-days saved a year (Grab Tech Blog, Jul 15, 2024).

    “Currently, all Azure OpenAI model deployments share the same quota under one account, which is set at 240K tokens per minute.”
    Hualin Liu, Grab engineer, lead author of the write-up, GrabGrab Tech Blog, , write-up

Choosing between OpenAI API and Azure OpenAI

Choose OpenAI API if

  • You build on features that run in OpenAI's own service, such as the Agents API (beta) and Secure MCP Tunnel.
  • You want a hard spend limit that stops traffic: OpenAI organizations and projects return 429 at a configured cap, while Azure budgets notify without stopping consumption.
  • You are starting fresh: OpenAI's Build tier allows 5,000 requests and 1,000,000 tokens per minute on GPT-6 Astra, against 1,000 and 1,000,000 on Azure's Tier 1 Global Standard.
  • Your agents depend on prompt caching through the Responses API: one developer measured about 96% cache hits on Azure Chat Completions and none on Azure's Responses API for GPT-5.6 Sol (Jul 12, 2026).

Choose Azure OpenAI if

  • Your organization buys through a Microsoft Enterprise Agreement, which Azure weighs when it assigns quota tiers.
  • You need Microsoft as the data processor, with prompts and completions kept out of OpenAI's own services.
  • You want provisioned throughput, with spillover to a standard deployment when the provisioned one is full.
  • You want a second route to the same models for outages and quota limits on OpenAI's endpoint, as DoorDash keeps behind its LLM Gateway.

Side by side

DimensionOpenAI APIAzure OpenAI
Monthly bill at one workloadGILD's calculation from published rates, fetched Oct 28, 2026: 2,000 tasks a day for 30 days, each with 2,000 new input tokens, 1,000 cache-write tokens, 7,000 cache-read tokens and 2,000 output tokens at GPT-6.1 Sol's US$2 and US$10. The long-prompt line sends 300,000 uncached input tokens and 2,000 output tokens per task at the long-context rates. Data Zone adds 10%. Taxes, reservations and negotiated discounts are excluded.US$1,632 a month on GPT-6.1 Sol with caching; US$73,800 if each task sends a 300,000-token prompt2US$1,632 a month on Global Standard, US$1,795.20 on Data Zone; US$73,800 on Global Standard for 300,000-token prompts1314
How you buy itFrom OpenAI, per token on prepaid credits, with Scale Tier and Reserved Tier for committed capacityOn an Azure subscription, as Azure OpenAI in Microsoft Foundry Models, with monthly and annual reservations for provisioned throughput
GPT-6 list pricesShort-context prices in US dollars per million input / output tokens; Azure prices from the Azure Retail Prices API for East US 2.GPT-6 Astra US$10 / US$50; GPT-6.1 Sol US$2 / US$10; GPT-6 Luna US$0.10 / US$0.50The same on Global Standard; Data Zone 10% higher (GPT-6 Astra US$11 / US$55)
Cached inputReads 0.1x the input rate (0.05x on GPT-6.1 Sol and GPT-6.1 Terra); 30-minute cache writes 1.25x and 12-hour writes 2.5xMetered at the same multipliers: GPT-6 Astra cached input US$1 and cache writes US$12.50 per million tokens on Global Standard
Discounted processingBatch and Flex at 50% off; Fast mode at 2x StandardBatch at 50% off Global Standard; Flex meters at 50% off on GPT-6 Global deployments
Starting throughputBuild tier: 5,000 RPM and 1,000,000 TPM on GPT-6 Astra, Sol, and Terra; 2,000,000 TPM on LunaTier 1: 1,000 RPM and 1,000,000 TPM per GPT-6 model on Global Standard; 333 RPM and 333,000 TPM on Data Zone
How limits growTiers upgrade automatically with total credit purchases: Launch at US$100 and Grow at US$500Seven quota tiers upgrade automatically with consumption, payment history, and Enterprise Agreement status
Data processorOpenAI; abuse monitoring logs kept up to 30 days; Zero Data Retention or Modified Abuse Monitoring by approvalMicrosoft, in its Azure environment; prompts and completions are not available to OpenAI; customers may apply for modified abuse monitoring
Data locationRegional processing in the US, the EU, and the UAE and storage-only residency in seven more regions, at a 10% uplift on eligible modelsGlobal, Data Zone (US, EU, or Asia Pacific), and regional deployments; data at rest stays in the resource's geography
Spend controlsOrganization and project hard spend limits that return 429 at the cap, plus spend alertsCost Management budgets and alerts, which do not stop consumption
Measured success rate98% of requests completed without an upstream error on Requesty's gateway in April 2026178% in the same April 2026 data, the lowest of the major routes Requesty measured1
Incidents teams reportOne developer's Responses API timeouts rose from 1.28% to 13.64% of calls over April 16 to 19, 2026, with no incident on OpenAI's status page26Microsoft's post-incident review records about six hours of failures and HTTP 500 and 503 errors in Sweden Central on Sep 29, 2026; a former customer left after latency spikes up to about 3 minutes (Apr 29, 2026)2527
Prompt caching on ResponsesOn by default; reads cost 0.1x the input rate (0.05x on GPT-6.1 Sol and GPT-6.1 Terra). On GPT-5.6 and later, writes cost 1.25x for the default 30-minute cache or 2.5x for a 12-hour cache6One developer measured about 96% cache hits on Azure Chat Completions and 0 cached tokens on Azure's Responses API for GPT-5.6 Sol (Jul 12, 2026)24
Agent toolingAgents SDK, the Agents API (beta) running a managed Codex harness, and remote MCP servers in the Responses APIThe Responses API, with multi-agent and computer use on supported models

27 sources

  1. Requesty: Operational metrics per provider, April 2026 (May 9, 2026)Public sourcerequesty.ai
  2. OpenAI API docs: PricingOfficial documentationdevelopers.openai.com
  3. OpenAI API docs: ModelsOfficial documentationdevelopers.openai.com
  4. OpenAI API docs: Rate limitsOfficial documentationdevelopers.openai.com
  5. OpenAI API docs: Spend limitsOfficial documentationdevelopers.openai.com
  6. OpenAI API docs: Prompt cachingOfficial documentationdevelopers.openai.com
  7. OpenAI API docs: Batch APIOfficial documentationdevelopers.openai.com
  8. OpenAI API docs: Data controls in the OpenAI platformOfficial documentationdevelopers.openai.com
  9. OpenAI API docs: AgentsOfficial documentationdevelopers.openai.com
  10. OpenAI API docs: MCP serversOfficial documentationdevelopers.openai.com
  11. Microsoft Learn: Foundry Models sold by AzureOfficial documentationlearn.microsoft.com
  12. Azure OpenAI in Foundry ModelsOfficial documentationazure.microsoft.com
  13. Azure OpenAI Service pricingOfficial documentationazure.microsoft.com
  14. Azure Retail Prices API: Azure OpenAI GPT6 meters, East US 2Official documentationprices.azure.com
  15. Microsoft Learn: Deployment types for Microsoft Foundry ModelsOfficial documentationlearn.microsoft.com
  16. Microsoft Learn: Azure OpenAI quotas and limitsOfficial documentationlearn.microsoft.com
  17. Microsoft Learn: Provisioned throughputOfficial documentationlearn.microsoft.com
  18. Microsoft Learn: Manage traffic with spillover for provisioned deploymentsOfficial documentationlearn.microsoft.com
  19. Microsoft Learn: Data, privacy, and security for Models sold by AzureOfficial documentationlearn.microsoft.com
  20. Microsoft Learn: Create and manage budgetsOfficial documentationlearn.microsoft.com
  21. InfoQ: Building a GenAI Platform at DoorDash, talk by Swaroop Chitlur and Siddharth Kodwani (Oct 3, 2026)Public sourceinfoq.com
  22. Grab Engineering: LLM-powered data classification for data entities at scale (Jul 15, 2024)Public sourceengineering.grab.com
  23. Grab Engineering: Grab AI Gateway, connecting Grabbers to multiple GenAI providers (Feb 19, 2025)Public sourceengineering.grab.com
  24. OpenAI Developer Forum: Prompt caching on Azure GPT-5.6 Sol, 96% hits on Chat Completions and none on Responses (Jul 12, 2026)Public sourcecommunity.openai.com
  25. Hacker News comment on Azure OpenAI latency spikes (Apr 29, 2026)Public sourcenews.ycombinator.com
  26. OpenAI Developer Forum: Responses API timeouts on gpt-4.1-mini rising from 1.28% to 13.64% (Apr 19, 2026)Public sourcecommunity.openai.com
  27. Azure status history: post-incident review for Azure OpenAI in Sweden Central (Sep 29, 2026)Official documentationazure.status.microsoft
Read the research