amBrain

Azure OpenAI alternative: when the retirement clock and the PTU bill decide for you

Published Sep 23, 2026Facts checked: Sep 23, 2026

Short answer

Azure OpenAI resells OpenAI's models inside an Azure subscription at the same list price: GPT-5.6-terra at $2.00 input and $12.00 output per million tokens on Global Standard. That is the short-context meter — the same model's long-context meters are $4.00 and $18.00, so a price quoted without its context tier understates a long-context bill by 50 to 100%.[1] Stay if procurement and compliance already run through Azure, if processing has to stay inside a US or EU data zone, and if provisioned capacity stays busy. Look for an alternative when the 18-month retirement clock and automatic upgrades keep reopening finished work, when provisioned throughput bills by the hour whether it is used or not, or when data must never leave your own servers. The alternatives are OpenAI direct, which already sells a GPT-6 generation whose gpt-6-sol costs $10.00 per million output against gpt-5.6-terra's $12.00[2]; Amazon Bedrock; or an open-weight model you host yourself.

On this page

What Azure OpenAI costs and how its models age

Azure OpenAI is OpenAI's models sold under a Microsoft contract, inside an Azure subscription, on Microsoft's datacentres. Three things decide whether that wrapper is worth it: which meter your traffic lands on, whether you pay for capacity by the hour, and how long a model you have tested stays available.

Azure's retail price list for East US 2 carries GPT-5.6-terra at $2.00 per million input tokens and $12.00 per million output on Global Standard — the meters 5.6 terra ShortCo Inp Std Gl and 5.6 terra ShortCo Opt Std Gl. Those are the short-context meters. The same model's long-context meters, 5.6 terra LongCo Inp Std Gl and 5.6 terra LongCo Opt Std Gl, read $4.00 and $18.00. Data Zone Standard adds exactly 10% to the short-context pair: 5.6 terra ShortCo Inp Std DZ at $2.20 and 5.6 terra ShortCo Opt Std DZ at $13.20.[1] A budget built on the headline $2.00 and $12.00 without naming the context tier is 50 to 100% short on long-context traffic.

OpenAI's own list carries the same $2.00 input and $12.00 output for gpt-5.6-terra, with cached input at $0.20, cache writes at $2.50, the Batch API at exactly half ($1.00 and $6.00) and Fast mode at exactly double ($4.00 and $24.00). Above that model now sits a GPT-6 flagship table: gpt-6-astra at $10.00 and $50.00 per million, gpt-6-sol at $2.00 and $10.00, gpt-6-luna at $0.10 and $0.50.[2] gpt-6-sol costs the same per input token as gpt-5.6-terra and less per output token, $10.00 against $12.00, so anchoring a comparison on gpt-5.6-terra now anchors it on a model its own successor undercuts.

Azure has not finished publishing that generation. Its pricing page already lists GPT-6 Astra in the minimum-PTU table, but the model prices come with a notice: "Notice: GPT-6 Sol and Luna prices are currently in processing for publishing on this page. Please find the model pricing in the blog."[3] Until that lands, the price of the model you would deploy today has to be read somewhere other than the pricing page.

The Batch API "returns completions within 24 hours for a 50% discount on Global Standard Pricing". Provisioned throughput is a different meter: "You are charged an hourly rate per model regardless of usage", with "monthly and annual reservations available to reduce overall spend" — the page states no term commitment either way. Minimum sizes for GPT-5.6-terra at short context are 15 PTUs on Global, 15 on Data Zone and 50 on Regional.[3] The retail price list puts those units at $1.00 an hour for Provisioned Managed Global Unit, $1.10 for Provisioned Managed Data Zone Unit and $2.00 for Provisioned Managed Regional Unit.[4]

An Azure OpenAI model ages on a published clock: the "Retirement date (18 months out) is set programmatically" at launch, new customers can no longer deploy it at 12 months, and "At 18 months from launch, all inference returns 410 Gone". Notice for generally available models is "At least 60 days", and "Retirement dates aren't extendable". Standard deployments are upgraded for you, region by region (gpt-4o version 2024-05-13 moves to gpt-5.1 on 2026-10-01), while "Provisioned deployments are NOT auto-upgraded" and are migrated by hand.[5] The 18 months do not generalise across the catalogue: on the same page, generally available models from Anthropic, DeepSeek, Fireworks and Mistral follow a 12-month lifecycle.[5]

The platform around the models has been renamed twice, and Microsoft's own migration page traces the chain: "Azure AI Studio → Azure AI Foundry → to Microsoft Foundry (current)". On August 26, 2026 the azure-ai-inference package retired and the Assistants API sunset; the recommended client became the standard OpenAI() with a base_url in place of AzureOpenAI(), authenticating through DefaultAzureCredential.[6] Renames are not outages, but each one moves documentation, portal paths and SDK names under work that was finished.

Azure OpenAI and the alternatives compared

Four routes to the same model call, by what each publisher states
Azure OpenAIOpenAI API directAmazon BedrockSelf-hosted open-weight model (vLLM)
Code and data ownershipPrompts, completions, embeddings and training data "are NOT available to OpenAI or other providers of Models sold by Azure" and "are NOT used to train any generative AI foundation models without your permission or instruction"; "Your fine-tuned Models sold by Azure are available exclusively for your use"[7]Services Agreement §4.1: "Customer: (a) retains all ownership rights in Input; and (b) owns all Output."[9] API data is not used for training unless you opt in, and abuse-monitoring logs are kept up to 30 days[8]"With Amazon Bedrock, your content is not used to improve the base models and is not shared with any model providers."[10]Prompts, outputs and weights stay on your servers
Cost modelGlobal Standard at OpenAI list prices: $2.00 / $12.00 per million short context, $4.00 / $18.00 long context, Data Zone +10%[1]. Provisioned units from $1.00 an hour[4], minimum 15 on Global[3]gpt-5.6-terra $2.00 / $12.00 per million, Batch $1.00 / $6.00, Fast mode $4.00 / $24.00; gpt-6-sol $2.00 / $10.00; regional endpoints add 10% for models released on or after March 5, 2026[2]"OpenAI models in Amazon Bedrock are billed through AWS. Bedrock pricing in commercial regions matches OpenAI direct pricing for equivalent services."[2]GPU hours or owned hardware: Standard_NC40ads_H100_v5 on Azure is $6.98 an hour on demand in East US[13]
Time to launchA resource and a deployment, subject to region: "Not all model and version combinations are available in all regions"[5]An API keyAn AWS account and model accessvllm serve openai/gpt-oss-120b gives a first endpoint[15]; the production controls around it are the work
Lock-in: notice, clock, upgradesRetirement 18 months after launch, "Retirement dates aren't extendable", "At least 60 days" notice for GA models; Standard deployments auto-upgraded, "Provisioned deployments are NOT auto-upgraded"[5]. "Monthly and annual reservations available to reduce overall spend", with no term commitment stated either way[3]"Generally available models: At least 6 months." of notice[12]"There are two Legacy periods: 6 months and 45 days. Most models have a 6-month Legacy period."; "existing customers may lose access after 15 days of inactivity"; the dates "may differ from dates published by model providers"[11]No retirement clock: you upgrade when your evaluation says so
CustomisationFine-tuning, with "Your fine-tuned Models sold by Azure are available exclusively for your use"[7]Closing: "OpenAI is winding down the fine-tuning platform. The platform is no longer accessible to new users, but existing users of the fine-tuning platform will be able to create training jobs for the coming months. All fine-tuned models will remain available for inference until their base models are deprecated."[2]Per model and per providerQuantization, "Efficient multi-LoRA support" and "structured outputs using xgrammar or guidance"[16]
Compliance and data residencyGlobal means "any geography where the relevant model sold by Azure is deployed"; Data Zone keeps processing in the US or the EU; Regional covers up to 27 regions; data at rest stays in the geography you designate[7]Exactly ten residency regions: US, Europe (EEA + Switzerland), Australia, Canada, Japan, India, Singapore, South Korea, United Kingdom and UAE. Outside the US "you must be approved for abuse monitoring controls, and execute a Modified Retention amendment"[8]"Any customer content processed by Amazon Bedrock is encrypted and stored at rest in the AWS Region where you are using Amazon Bedrock."[10]Wherever your servers are
Latency and scaleProvisioned throughput reserves capacity; "Microsoft can limit new customers in specific regions"[5]Shared capacity, with Fast mode at exactly double the standard price: $4.00 and $24.00 per million[2]Provisioned Throughput for some models, but "You can't create a new Provisioned Throughput for models in the Legacy state."[11]Dedicated GPUs, with "Tensor, pipeline, data, expert, and context parallelism" across them[16]
Exit costLow in code: Microsoft's own mapping replaces AzureOpenAI() with OpenAI() plus a base_url and describes the result as "Azure-specific code eliminated"[6]. Re-testing every workload is the real costRe-test on the next modelRe-test on the next modelvLLM serves an OpenAI-compatible API, so the same client shape reaches a self-hosted endpoint[17]; hardware and the team that runs it are the cost

Stay on Azure OpenAI if…

  1. 1.Procurement and compliance already run through Azure. Data Zone deployments keep processing inside the US or the EU, and data at rest stays in the geography you designate.[7] Changing provider reopens reviews you have already closed.
  2. 2.Your load keeps provisioned capacity busy. "You are charged an hourly rate per model regardless of usage", which is a cost only when the units idle, and the page offers "monthly and annual reservations available to reduce overall spend".[3]
  3. 3.Your identity setup is Microsoft's. The current client authenticates through DefaultAzureCredential,[6] and "Your fine-tuned Models sold by Azure are available exclusively for your use".[7]

Consider leaving Azure OpenAI if…

  1. 1.Symptom: every retirement reopens tested work. Prompts, parsers and evaluation sets are redone on Microsoft's calendar — the "Retirement date (18 months out) is set programmatically", "Retirement dates aren't extendable", and Standard deployments are upgraded for you, as gpt-4o 2024-05-13 moves to gpt-5.1 on 2026-10-01.[5] The model generation moves faster than the clock: OpenAI already lists a GPT-6 family[2] while Azure's pricing page still says Sol and Luna prices are "currently in processing for publishing on this page".[3]
  2. 2.Symptom: the provisioned line is the same size in a quiet month. At the published hourly rate, the 15-PTU Global minimum costs 15 × $1.00 × 730 hours = $10,950 a month, and the 50-PTU Regional minimum 50 × $2.00 × 730 = $73,000.[3][4] That is arithmetic on list prices, not a figure Microsoft quotes.
  3. 3.Symptom: a contract or a regulator says "our infrastructure only". Azure OpenAI runs in Microsoft's datacentres, and a Global deployment may process a prompt in "any geography where the relevant model sold by Azure is deployed".[7]

A migration path in four steps

  1. 1.Inventory deployments and dates. List every deployment, model version and SKU with its lifecycleStatus and deprecationDate from the Models API,[5] and mark which of them a monthly or an annual reservation covers.[3]
  2. 2.Move clients to the standard OpenAI client. Microsoft's own mapping replaces AzureOpenAI() with OpenAI() and a base_url, and calls the result "Azure-specific code eliminated".[6] The same client shape then reaches a self-hosted endpoint, because vLLM serves an OpenAI-compatible API.[17]
  3. 3.Evaluate the target on your own test set. Run shadow traffic per workload and compare against the answers you accept; the AI pilot that never reached production explains why the test set comes first, and OpenAI API vs a self-hosted LLM puts the two targets side by side.
  4. 4.Move workload by workload. Shift traffic where the target passes, then delete the Azure deployments once nothing routes to them. Keeping one provider as a fallback is a routing rule, covered in LLM in a closed perimeter.

What drives the cost of hosting your own

Compare like with like. The 15-PTU Global minimum at the published hourly rate is about $10,950 a month — 15 × $1.00 × 730.[3][4] That buys roughly two on-demand H100 machines on Azure: Standard_NC40ads_H100_v5 is $6.98 an hour in East US, and 2 × $6.98 × 730 is $10,190.80 a month.[13] The capacities are not equivalent. The PTUs serve GPT-5.6; the GPUs serve an open-weight model such as gpt-oss-120b, which is Apache 2.0 and is quantised with the aim of "making gpt-oss-120b run on a single 80GB GPU".[15] Score that model on your evaluation set before comparing invoices: the price says nothing about whether it is good enough.

These are the cost drivers, not a price.

What moves the bill when the model runs on your own hardware
Cost driverWhat drives itPublic reference point
GPU capacityModel size, context length, concurrencyStandard_NC40ads_H100_v5, Consumption Linux, East US: $6.98 an hour[13]. Microsoft's size page describes the series as "powered by NVIDIA H100 NVL GPU", Accelerators (Qty.) 1, Accelerator-Memory (GB) 94[14]
Model fit and licenceParameters, quantisation, commercial termsgpt-oss-120b is Apache 2.0 and quantised with the aim of "making gpt-oss-120b run on a single 80GB GPU", while "the gpt-oss-20b model run within 16GB of memory"[15]
RedundancyA second machine so one failure does not stop the serviceDoubles the GPU line: 2 × $6.98 × 730 = $10,190.80 a month[13]
Security layerAuthentication, network isolation and logging in front of the servervLLM's --api-key "only authenticates requests to endpoints under the /v1, /v2, and /inference path prefixes", /invocations is not authenticated, and the docs say "Do not rely on --api-key alone to secure vLLM"[17]
EvaluationA test set per workload, rerun on every model changeNo public reference point; scoped per project
OperationsUpgrades, monitoring, on-callNo public reference point; scoped per project

Frequently asked questions

On Global Standard, no: GPT-5.6-terra lists at $2.00 input and $12.00 output per million on both — and on Azure that is the short-context meter, with long context at $4.00 and $18.00.[1][2] Data Zone Standard costs exactly 10% more on Azure ($2.20 and $13.20), and OpenAI's own regional endpoints add 10% for models released on or after March 5, 2026.[1][2]

For an Azure OpenAI model, 18 months after launch: the "Retirement date (18 months out) is set programmatically" and is visible through the Models API. Generally available models get "At least 60 days" notice, and "Retirement dates aren't extendable".[5] The figure does not carry across the catalogue — GA models from Anthropic, DeepSeek, Fireworks and Mistral on the same platform follow a 12-month lifecycle.[5]

Standard deployments are upgraded automatically to the replacement model (gpt-4o 2024-05-13 to gpt-5.1 on 2026-10-01), while "Provisioned deployments are NOT auto-upgraded" and have to be migrated by hand. After the date, "all inference returns 410 Gone".[5]

No. Microsoft states that prompts, completions, embeddings and training data "are NOT available to OpenAI or other providers of Models sold by Azure", and that they "are NOT used to train any generative AI foundation models without your permission or instruction".[7]

OpenAI's Services Agreement says it in §4.1: "Customer: (a) retains all ownership rights in Input; and (b) owns all Output."[9] Separately, the data-controls page states that API data is not used for training unless you opt in and that abuse-monitoring logs are kept up to 30 days.[8]

Not as a new customer. OpenAI's pricing page states: "OpenAI is winding down the fine-tuning platform. The platform is no longer accessible to new users, but existing users of the fine-tuning platform will be able to create training jobs for the coming months. All fine-tuned models will remain available for inference until their base models are deprecated."[2] On Azure, fine-tuning is still offered, and "Your fine-tuned Models sold by Azure are available exclusively for your use".[7]

OpenAI's open-weight models, yes: gpt-oss-120b and gpt-oss-20b are Apache 2.0, quantised with the aim of "making gpt-oss-120b run on a single 80GB GPU" and of letting "the gpt-oss-20b model run within 16GB of memory".[15] The GPT-5.6 and GPT-6 families are sold through APIs — OpenAI and Azure both list them,[2][1] and OpenAI states that "OpenAI models in Amazon Bedrock are billed through AWS."[2] Which route fits is the question behind OpenAI API vs a self-hosted LLM and vLLM vs Ollama.

About amBrain

  • We have taken an LLM integration to production inside a client's FinTech perimeter: extracting and normalising unstructured broker and venue notices — corporate actions, instrument and margin changes — into structured records the trading system consumes.
  • Three formats: full delivery, a dedicated team, or engineers embedded in your team.
  • The client keeps full ownership of the product and the code, except our reusable components.

Disclosure: this page is published by amBrain. The last column of the comparison table is the kind of work amBrain does. Every vendor fact on this page links to its source and carries the date that source was read; model prices move often, and every figure here was read on 2026-09-23. No vendor named here reviewed or endorsed this page. Azure, Microsoft Foundry, OpenAI, GPT, Amazon Bedrock, AWS, NVIDIA and vLLM are trademarks of their owners and are used only to identify the products discussed.

Sources

  1. [1]Microsoft Azure, Retail Prices API, Foundry Models meters, East US 2 (JSON payload: the short- and long-context GPT-5.6-terra meters for Global Standard and Data Zone Standard, with their unit prices). Vendor's own page · checked Sep 23, 2026
  2. [2]OpenAI, Pricing (GPT-6 and GPT-5.6 token tables, Batch and Fast mode, the regional uplift, the Bedrock billing note and the Finetuning section). Vendor's own page · checked Sep 23, 2026
  3. [3]Microsoft Azure, Azure OpenAI Service - Pricing (deployment types, the Batch discount wording, provisioned billing, reservations, the minimum-PTU table and the GPT-6 pricing notice). Vendor's own page · checked Sep 23, 2026
  4. [4]Microsoft Azure, Retail Prices API, Azure OpenAI provisioned meters, East US 2 (JSON payload: Provisioned Managed Global, Data Zone and Regional units with their hourly prices). Vendor's own page · checked Sep 23, 2026
  5. [5]Microsoft Learn, Foundry Models lifecycle and support policy (retirement dates, notice periods, automatic upgrades, regional availability; the 18-month clock on this page belongs to Azure OpenAI models, while GA models from Anthropic, DeepSeek, Fireworks and Mistral follow a 12-month lifecycle). Official documentation · checked Sep 23, 2026
  6. [6]Microsoft Learn, Migrate from the Foundry (classic) portal (the platform naming chain, the August 26, 2026 retirement dates and the SDK mapping from AzureOpenAI() to OpenAI(); the page covers migration to the OpenAI client, not portability to any other endpoint). Official documentation · checked Sep 23, 2026
  7. [7]Microsoft Learn, Data, privacy, and security for Foundry Models sold by Azure in Microsoft Foundry. Official documentation · checked Sep 23, 2026
  8. [8]OpenAI, Data controls in the OpenAI platform (training opt-in, abuse-monitoring retention and the residency region list; the page covers retention, residency and training only, and says nothing about ownership). Official documentation · checked Sep 23, 2026
  9. [9]OpenAI, OpenAI Services Agreement (version ONLINE v.010126, §4.1). Vendor's published terms · checked Sep 23, 2026
  10. [10]Amazon Web Services, Amazon Bedrock FAQs (how customer content is used and where it is stored at rest). Vendor's own page · checked Sep 23, 2026
  11. [11]Amazon Web Services, Model lifecycle, Amazon Bedrock User Guide (Legacy periods, the inactivity rule and the Provisioned Throughput restriction; the page scopes itself to models launched on Bedrock on or after September 7, 2026). Official documentation · checked Sep 23, 2026
  12. [12]OpenAI, Deprecations (the notice period OpenAI commits to for generally available models). Official documentation · checked Sep 23, 2026
  13. [13]Microsoft Azure, Retail Prices API, Virtual Machines H100 sizes, East US (JSON payload: Standard_NC40ads_H100_v5, Consumption Linux, $6.98 per hour; the payload carries no GPU count and no memory size). Vendor's own page · checked Sep 23, 2026
  14. [14]Microsoft Learn, NCads_H100_v5 size series (the hardware behind the SKU: accelerator model, quantity and memory). Official documentation · checked Sep 23, 2026
  15. [15]OpenAI, openai/gpt-oss-120b model card (licence, quantisation targets and the first serving command). Vendor's own page · checked Sep 23, 2026
  16. [16]vLLM, Source repository and README. Vendor's own page · checked Sep 23, 2026
  17. [17]vLLM, OpenAI-Compatible Server, serving documentation (the OpenAI-compatible surface and the limits of the built-in API key check). Official documentation · checked Sep 23, 2026

Related

Free project intro call, 30 minutes.

We reply within 24 hours.

Book a Call