Token price will not settle this decision: GPT-6.1 Sol costs $2.00 per million input tokens and $10.00 per million output tokens on OpenAI's API, Azure OpenAI and Amazon Bedrock. All three charge 10% more to keep processing inside a region.
If your company has a Microsoft agreement, buy GPT-6 through Azure OpenAI on Data Zone Standard pay-as-you-go. You get EU or US processing, Entra ID, private networking, a 99.9% uptime SLA and spend that counts against your Azure commitment, at the same rate as OpenAI. Open a second, direct OpenAI account for the features Azure does not carry yet, chiefly Ultrafast and OpenAI's hosted Agents API. Do not buy Azure provisioned throughput to save money, because for the GPT-6 family Microsoft prices it at parity with pay-as-you-go at 100% utilization. Bedrock is the weakest of the three for this buyer: it has no EU-bound route for these models and no reserved capacity.
| OpenAI API (direct) | Azure OpenAI in Foundry | OpenAI on Amazon Bedrock | |
|---|---|---|---|
| GPT-6.1 Sol, global, per 1M tokens (input / cached / output) | $2.00 / $0.10 / $10.00 | $2.00 / $0.10 / $10.00 | $2.00 / $0.10 / $10.00 |
| In-region processing | +10%, US, EU and UAE | +10%, US, EU and Asia Pacific Data Zones | +10%, US only |
| Reserved capacity | Reserved Tier, contact sales | PTU, $260 per PTU-month (Global) | Not offered for GPT-6.1 Sol |
| Speed tiers on GPT-6.1 Sol | Fast mode, Ultrafast | None found | Ultrafast |
| Published uptime SLA | None on Standard; 99.9% on Fast mode | 99.9% | 99.9% per region |
| Workload identity | API keys or workload identity federation | Entra ID and managed identities | IAM and Bedrock API keys |
| Counts against a Microsoft MACC | No | Yes, per Redress | No |
| Our verdict | Second account, for features | Default for Microsoft shops | Skip unless AWS-only and US-only |
Prices were checked on November 10, 2026 on OpenAI's pricing page, Microsoft's Azure Retail Prices API (the Azure OpenAI pricing page renders its numbers in the browser) and AWS's GPT-6.1 Sol model card. Enterprise discounts are private on all three, so treat these as list prices.
Is GPT-6 Cheaper on Azure OpenAI or on the OpenAI API?
Neither is cheaper at list price. Every GPT-6 and GPT-6.1 rate we compared is identical on the global tier, and the in-region premium is also identical at 10%.
On OpenAI's pricing page, GPT-6.1 Sol Standard costs $2.00 input, $0.10 cached input and $10.00 output per million tokens, and GPT-6 Astra costs $10.00, $1.00 and $50.00. Batch and Flex are half that. The page says regional processing endpoints carry "a 10% uplift for models released on or after March 5, 2026."
Azure's retail price list for East US 2 shows GPT-6.1 Sol Global Standard at $2.00 input, $0.10 cached input and $10.00 output, and Data Zone Standard at $2.20, $0.11 and $11.00, according to Microsoft's Retail Prices API. GPT-6 Astra matches as well: $10.00 and $50.00 on Global, $11.00 and $55.00 on Data Zone. Long-context rates (prompts above 272K tokens) match too, at $4.00 and $15.00 for GPT-6.1 Sol on both.
AWS states the parity outright. Its GPT-6.1 Sol model card says "Global CRIS rates match OpenAI first-party pricing for the same service tier," and in-Region or US cross-Region access costs 10% more.
Normalize to one workload to see the result. Take GPT-6.1 Sol at 1 billion fresh input tokens, 1 billion cached input tokens and 200 million output tokens a month, all short context. Cache writes bill at $2.50 per million on all three routes, so we leave them out.
| Route | Global processing | In-region processing |
|---|---|---|
| OpenAI API | $4,100 | $4,510 (US, EU or UAE) |
| Azure OpenAI | $4,100 | $4,510 (Data Zone) |
| Bedrock | $4,100 | $4,510 (US only) |
What Azure Provisioned Throughput Actually Buys You
Azure provisioned throughput (PTU) buys reserved capacity and a published latency target; for GPT-6 it does not buy a discount unless you commit for a year. A PTU is a unit of model capacity you rent by the hour, month or year, sized in tokens per minute.
The retail price list puts Global provisioned at $1.00 per PTU-hour, Data Zone at $1.10 and Regional at $2.00. Monthly reservations cost $260 per PTU (Global) or $286 (Data Zone and Regional), and one-year reservations cost $2,652 or $2,916, per the Retail Prices API.
Microsoft's own PTU sizing page does the math for GPT-6 Sol. One PTU delivers 3,000 normalized tokens a minute, or 129.6 million a month at full load, which costs $259.20 at pay-as-you-go rates against $260 for the PTU. Microsoft's conclusion: "PTU and pay-as-you-go costs are approximately equal at full utilization." A one-year reservation works out to $221 a month, about 15% below pay-as-you-go, and only if you keep the PTUs fully busy.
Apply that to the workload above. GPT-6.1 Sol's normalized-token weights are 1.0 for input, 0.05 for cached input and 5.0 for output, so the month comes to 2.05 billion normalized tokens, or 15.8 PTUs at a perfectly flat load. Real traffic peaks. At a 3:1 peak-to-average ratio you need 50 PTUs, which costs $13,000 a month on Global monthly reservations against $4,100 on pay-as-you-go.
Buy PTU for the latency target, which is 99% of 5-minute windows above 50 tokens per second for GPT-6.1 Sol on the sizing page, and for capacity you cannot lose during someone else's launch. Our PTU break-even analysis has the utilization math for annual terms.
OpenAI's equivalent is Reserved Tier, sold to Enterprise customers through sales, bought per model in dollars per minute, with usage measured at pay-as-you-go list prices and discount handling explained in an FAQ. Bedrock lists the Reserved tier as unsupported for GPT-6.1 Sol and GPT-6 Astra on their model cards.
Which Route Has the Better SLA?
Azure has the stronger published paper for a GPT workload: a 99.9% uptime SLA on Microsoft's product page plus per-model latency targets on provisioned deployments. OpenAI publishes no uptime SLA for Standard pay-as-you-go traffic, according to a DevHelm summary of its terms. OpenAI's Fast mode page lists 99.9% uptime.
Bedrock's SLA, last updated October 4, 2023, commits to 99.9% per region and pays 10% below 99.9%, 25% below 99.0% and 100% below 95.0%. It has no latency term, and it does not name third-party models either way, so ask AWS to confirm in writing that OpenAI models are covered.
The exclusions matter more than the headline. Redress Compliance reads Microsoft's SLA as counting a 429 throttle as available and reports credits of 10% below 99.9% and 25% below 99%. Microsoft's latency targets are a p50 per 5-minute window, so your slowest requests have no floor on any route. Our SLA comparison covers the credit mechanics in full.
How Far Behind OpenAI Does Azure Run in 2026?
For GPT-6 models the lag is now days, and for some features it is open-ended. Model launches reach Azure almost at once, while the service tiers and hosted products around them often do not reach Azure at all.
OpenAI's changelog dates GPT-6 Astra to September 3, GPT-6 Sol and Luna to September 22, GPT-6.1 Sol to September 29 and GPT-6.1 Terra to October 26. Microsoft's model list carries the same version dates for all five. Astra arrived in Foundry through a Limited Access Program by September 4, and StorageReview reported it available across Global and US Data Zone the next day. Bedrock's model card gives Astra a launch date of September 8.
The gaps sit around the models:
- OpenAI added Ultrafast for GPT-6 Astra on September 29 and for GPT-6.1 Sol on October 8, at six times the Standard price. Bedrock sells it for both models. We found no Ultrafast option in Foundry's model documentation or Azure's retail meters on November 10.
- Azure's priority processing page lists gpt-6-sol and no other GPT-6 or GPT-6.1 model, and it excludes EU Data Zone deployments. OpenAI's pricing page lists Fast mode for both GPT-6 Astra and GPT-6.1 Sol.
- Microsoft's region availability tables show no GPT-6 family model on Regional Standard, and no GPT-6.1 Sol on Regional Provisioned or Global Batch. GPT-6 Astra is missing from the Asia Pacific Data Zone.
- OpenAI's Agents API runs computer use in an OpenAI-hosted browser, added September 29 per the changelog. It runs on OpenAI's infrastructure, so an Azure deployment does not include it.
The Regional Standard gap hurts most in countries without a Data Zone. In a July 2026 Microsoft Q&A thread, a customer needing Canadian residency found no GPT-5.1 on Regional Standard in either Canadian region, and a Microsoft moderator replied that Microsoft "has not published a roadmap or timeline" for one. GPT-6 has the same gap.
Data Handling: Retention, ZDR and Where Prompts Are Processed
Azure keeps OpenAI out of the data path entirely; OpenAI's direct API gives you more processing regions but makes you sign a retention amendment for most of them.
Microsoft's data privacy page says prompts and completions "are NOT available to OpenAI" and that the models "do NOT interact with any services operated by" OpenAI. A Data Zone deployment in an EU member state may process prompts "in that or any other European Union Member Nation," and stored data, including the abuse-monitoring store, stays in the resource's geography. Modified abuse monitoring, which turns off storage and human review, is an application open to customers managed by a Microsoft account team or an eligible program.
OpenAI's data controls guide keeps abuse-monitoring logs "for up to 30 days" by default. Zero data retention and modified abuse monitoring are "subject to prior approval by OpenAI," and several endpoints (files, batches, vector stores, Conversations, agents) are never ZDR-eligible. Regional processing runs in the US, Europe and the UAE. Australia, Canada, Japan, India, Singapore, South Korea and the UK are storage-only, so inference can happen elsewhere. Outside the US you must be approved for abuse-monitoring controls and sign a Modified Retention amendment, and EU regional processing also requires ZDR, MAM or one of OpenAI's retention programs.
Bedrock's route for these models is US-only for in-region work. The GPT-6 Astra model card offers European regions such as Frankfurt, Ireland and Paris only through Global cross-Region inference, which "routes requests anywhere in the world." Its strength is retention control: a none data-retention mode that an SCP can enforce across the organization. Our ZDR comparison walks through each vendor's exclusions.
Identity, Networking and Your Azure Commitment
Azure fits a Microsoft security stack with no extra work. Microsoft's keyless connection guidance recommends Entra ID with managed identities for Azure-hosted apps, so you assign an RBAC role and stop handing out keys. Private endpoints carry traffic over Azure Private Link and let you block the public endpoint.
The OpenAI API is no longer keys-only. Its workload identity federation exchanges a short-lived token from Azure, AWS, Google Cloud, Kubernetes, GitHub Actions or SPIFFE for a short-lived OpenAI token. It does not give you Entra conditional access or Azure RBAC on the deployment, and we found no private-network option in OpenAI's public documentation. Bedrock supports AWS PrivateLink on both the bedrock-runtime and bedrock-mantle endpoints.
Billing is the deciding line for most Microsoft shops. Redress's negotiation guide says "Azure OpenAI consumption counts toward the Microsoft Azure Consumption Commitment on both reserved PTU and pay as you go," while some Marketplace-billed third-party models may not. An invoice from OpenAI is not an Azure meter and cannot draw down a MACC. If your MACC is going to under-run, a GPT dollar spent direct is new spend, while the same dollar on Azure reduces a shortfall you would otherwise owe. Microsoft's MACC tracking page shows which services decrement your commitment; check it before you sign.
The Case Against Azure, and Who Should Stay Direct
Azure's weak points are quota, access to new tiers, and deployment types that never reach some regions.
Microsoft's quota documentation uses subscription tiers. Microsoft's model page says some tiers must file quota requests to deploy GPT-6 models, and only Tier 5 and Tier 6 subscriptions have that quota by default. Free-tier subscriptions cannot request increases at all, and a team that exceeds its tier can see latency "more than two times higher." On the direct API, OpenAI's GPT-6.1 Sol model page starts its lowest listed tier at 1 million tokens per minute.
Skip Azure OpenAI if you need Ultrafast, Fast mode on GPT-6 Astra or GPT-6.1 Sol, OpenAI's hosted browser agents, or single-country processing for GPT-6 outside the US, EU and Asia Pacific zones. Skip it if you buy self-serve with no account team, because you cannot apply for modified abuse monitoring.
Skip the direct OpenAI API as your main route if you have a MACC to spend, if your customers require a published upstream uptime SLA on standard traffic, or if your security team requires a private network path. Skip it for EU processing if legal will not sign a retention amendment.
Skip Bedrock if any GPT-6 traffic must stay in the EU, if you want reserved capacity, or if you have not confirmed how Marketplace-billed models count against your AWS commitment. Bedrock lists a Marketplace product ID on each OpenAI model card, so that answer sits in your EDP terms. Bedrock makes sense only for an AWS-committed team whose workload can run in the US and that wants Ultrafast without an OpenAI contract.
How to Decide: The Criteria That Predict Regret
Four questions decide this, in order:
- Do you have a MACC that will under-run? If yes, Azure wins before anything else is weighed.
- Where must inference run? For the EU or the US, use Azure Data Zone or OpenAI regional processing. For the UAE, use OpenAI direct. If it must stay in one EU country or in Canada, none of the three routes offers GPT-6 pay-as-you-go processing there today, so check Azure Regional Provisioned availability for your model and region, or wait.
- Do you need a speed tier on the newest models? Ultrafast on GPT-6.1 Sol or GPT-6 Astra means OpenAI direct or Bedrock.
- Can you get quota? Check your Azure tier before you set a launch date.
The answer changes if Microsoft adds Ultrafast or Regional Standard for GPT-6, if Bedrock adds an EU geographic profile, or if OpenAI publishes a private networking option.
This Week:
- Pull your Azure quota tier and your MACC drawdown forecast, and put both in front of whoever owns the GPT budget.
- Price your real token mix on each route with your cached-input share, using the $2.00 / $0.10 / $10.00 GPT-6.1 Sol rates above.
This Month:
- Deploy GPT-6.1 Sol on Data Zone Standard, move one production workload, and log tokens per second and 429 rates for two weeks.
- Open a direct OpenAI organization with workload identity federation for the workloads that need Ultrafast or hosted agents, and route them through your gateway so both bills stay visible.
Before Q1 Close:
- Ask Microsoft for the current SLA document and the exact credit tiers, and ask OpenAI sales for Reserved Tier discounts in writing.
- Buy PTU only for a customer-facing path that needs the latency target, sized at your measured peak, never to cut cost.
The Bottom Line
Three vendors now sell the same OpenAI models at the same token price, much as the big clouds all resell the same database engines and differ on support, networking and billing. For a company with a Microsoft agreement, Azure's terms are the best of the three: commit drawdown, Entra ID, private endpoints and a published SLA. Keep a direct OpenAI account for the features Azure lacks, and re-run this comparison when Microsoft adds Ultrafast.
Continue Reading
- OpenAI, Azure, Bedrock, Vertex and Anthropic SLAs Compared
- Annual Azure and Google AI reservations break even at 84% and 75%
- Use Bedrock's None Mode to Enforce Zero Data Retention
- LLM Data Residency: Which Providers Actually Keep Data In Region
- Bedrock vs Vertex AI vs Foundry: Your Cloud Commit Decides
- GPT-6 Astra Ultrafast Costs 6x Standard and Excludes EU Workloads
- Google Cloud Lets Codex Users Run OpenAI Models on Its Platform