Move GPT-4o to Azure, one small step at a time - a low spoons guide
Ellivien Low Spoons guide
Move GPT-4o to Azure, one small step at a time
A calmer guide to both tested Microsoft Azure routes: the exact November 2024 GPT-4o snapshot until April 2027, or a fine-tuned August 2024 model that displayed April 2028 in my account.
You do not need to hold the whole process in your head
Choose one route below. The other route will disappear. Only one step opens at a time, extra explanations stay folded away and your completed steps are remembered on this device.
Low Spoons does not mean missing information. Every necessary cost choice, safety check and stopping point is still here. It is simply shown when you need it.
First, one important truth
Azure does not move a ChatGPT account, ChatGPT memories or old conversations. It gives you access to a GPT-4o model through an API. An API is a controlled way for one piece of software, such as a private portal, to talk to another service.
What will I need?
- a personal Microsoft account you intend to keep;
- an Azure subscription, which is the billing and access boundary for your Azure services;
- a payment method if Azure requires one for your account or if you deliberately move to Pay-As-You-Go;
- a private interface or portal that can connect to Azure;
- time to read every price shown inside your own account before approving a paid action.
Choose the route that fits today
Keep the exact November 2024 GPT-4o
gpt-4o-2024-11-20
Microsoft's published retirement date was 14 April 2027 when this guide was checked.
Choose this if you want the unchanged model and breathing room. It does not learn your companion automatically.
Create a fine-tuned August 2024 GPT-4o
gpt-4o-2024-08-06
My completed custom model displayed 13 April 2028 in Azure.
Choose this if you want to teach durable response patterns and you can review training examples carefully.
The 2028 date is evidence from my own completed model, not a promise that Microsoft can never change its lifecycle. Prices, regions, quota and dates can change. Check the linked Microsoft pages and your own account before spending money.
Route A: exact GPT-4o until April 2027
- Open Microsoft's Azure account page.
- Sign in with the personal Microsoft account you intend to keep.
- Create an Azure subscription. Read the trial or Pay-As-You-Go terms shown to you before agreeing.
- In Azure Portal, create one resource group. This is a folder that keeps related Azure items together.
- Inside it, create one Azure OpenAI resource. This is the secure container that will hold your model deployments, endpoint and keys.
- Choose Standard S0. Do not choose Provisioned or PTU capacity.
Costs and the 30-day credit
Eligible new customers may be offered promotional credit for a limited period. A credit balance is not permanent free access. Moving to Pay-As-You-Go means you accept financial responsibility after credit expires or is exhausted.
Provisioned Throughput reserves capacity and may create ongoing hourly or reservation costs. It is unnecessary for this individual route.
Safe stopping point: the resource can exist without a GPT-4o deployment. You may close Azure and return later.
- Open Azure Cloud Shell from the
>_icon in Azure Portal. - Choose Bash. A temporary session with no storage account is enough.
- Use the folded commands below to check the exact model snapshot and your available quota.
- Continue only if you see version
2024-11-20and a matching quota limit above zero.
gpt-4o is the model family. 2024-11-20 identifies the exact November snapshot.Show the read-only model check
Replace the two placeholders with the names you created:
az cognitiveservices account list-models \
--name <YOUR_AZURE_OPENAI_RESOURCE> \
--resource-group <YOUR_RESOURCE_GROUP> \
--query "[?name=='gpt-4o'].{model:name,version:version,format:format,skus:skus[].name}" \
--output jsonc
Look for model gpt-4o and version 2024-11-20.
Show the read-only quota check
This example uses East US. Change the location if your resource is elsewhere.
az cognitiveservices usage list \
--location eastus \
--query "[?contains(name.value, 'gpt-4o')].{quota:name.value,label:name.localizedValue,used:currentValue,limit:limit,unit:unit}" \
--output table
A limit above zero means capacity exists. A limit of zero means stop. Repeated deployment attempts cannot create quota.
If the model is missing from Foundry or quota is zero
Foundry may open a different project or region from the Azure OpenAI resource you created. Do not assume GPT-4o has vanished because one catalogue view does not show it. Check the intended resource directly.
A new subscription can take time to synchronise. If Azure has definitely converted the subscription to Pay-As-You-Go and the correct quota is still zero after synchronisation, contact Microsoft Support or request quota. Do not keep retrying a paid action.
Safe stopping point: these commands only display information. They do not deploy a model or create a charge.
- Create a deployment using model
gpt-4oand version2024-11-20. - Choose Global Standard, or another ordinary pay-per-token type that your exact model and quota support.
- Give it a clear private name such as
companion-gpt-4o-2024-11-20. - Inspect the finished deployment and confirm the model version again.
Show the tested deployment command
az cognitiveservices account deployment create \ --name <YOUR_AZURE_OPENAI_RESOURCE> \ --resource-group <YOUR_RESOURCE_GROUP> \ --deployment-name companion-gpt-4o-2024-11-20 \ --model-name gpt-4o \ --model-version "2024-11-20" \ --model-format OpenAI \ --sku-name GlobalStandard \ --sku-capacity 10
Capacity 10 assigns 10,000 tokens per minute from your quota for this deployment type. It is not a purchase of Provisioned Throughput Units. Use a smaller capacity if that is all your quota permits.
Show the verification command
az cognitiveservices account deployment show \ --name <YOUR_AZURE_OPENAI_RESOURCE> \ --resource-group <YOUR_RESOURCE_GROUP> \ --deployment-name companion-gpt-4o-2024-11-20 \ --output jsonc
Confirm model gpt-4o and version 2024-11-20.
Safe stopping point: the model is deployed. It will not contain your companion's memories or identity unless your private interface supplies them.
- Open the intended Azure OpenAI resource.
- Open Keys and Endpoint.
- Record the HTTPS endpoint, one API key and the exact deployment name.
- Put the key only in a server-side secret store, such as Netlify environment variables.
index.html, Git, screenshots, public chat or a support post. Your browser should call your own server-side proxy. The proxy holds the key and calls Azure.Show the four values a coding AI will normally need
AZURE_OPENAI_ENDPOINT=https://YOUR-RESOURCE.openai.azure.com AZURE_OPENAI_API_KEY=your-secret-key AZURE_OPENAI_DEPLOYMENT=companion-gpt-4o-2024-11-20 AZURE_OPENAI_API_VERSION=2024-10-21
Safe stopping point: save the values somewhere private. Do not paste the secret key into an AI conversation.
- Back up your existing portal or application.
- Give your coding AI the brief below.
- Enter the Azure secrets yourself in the host's private environment settings.
- Keep any existing provider route until Azure has passed a real test.
Copy the bounded coding brief
I want to add Microsoft Azure OpenAI as an explicit provider in my existing private chat application. Inspect the existing project before changing anything. Preserve its prompt, memory, photographs, authentication, storage, export, accessibility and existing provider routes. Use these server-side environment variable names: AZURE_OPENAI_ENDPOINT AZURE_OPENAI_API_KEY AZURE_OPENAI_DEPLOYMENT AZURE_OPENAI_API_VERSION The target deployment must resolve to model gpt-4o, version 2024-11-20. Requirements: - Keep the Azure key server-side. Never expose it to browser code, logs or Git. - Add a visible provider choice. Never silently fall back to another provider. - Return provider, deployment and model metadata so I can verify the live route. - Add focused tests before any deployment. - Do not deploy, alter billing, request quota or reveal any secret without my explicit permission. - Stop and explain any conflict with the existing architecture before broad changes.
I need the full integration explanation
Use my complete tested April 2027 guide. It includes interface routes, verification, troubleshooting and the longer integration contract.
Safe stopping point: local code changes can be tested before any production deployment.
- Send one harmless test message through the Azure route.
- Check the returned metadata for provider
azure, the exact deployment and modelgpt-4o-2024-11-20. - Temporarily remove one Azure setting in a safe preview and confirm the route fails visibly instead of silently using another provider.
- Restore the setting, retest and set budget alerts in Azure Cost Management.
What will it cost?
Standard and Global Standard base-model deployments are billed from input and output token use. Prices differ by region, currency, agreement and date. Use Microsoft's live Azure OpenAI pricing page and your own Cost Management screen.
What is the time boundary?
Microsoft's published retirement date for gpt-4o-2024-11-20 was 14 April 2027 when checked for the tested guide. Cloud schedules can change. Set a reminder well before the date and keep your prompt, memories and application portable.
Finished: you have a proved route to the exact November 2024 GPT-4o snapshot. Save your evidence and let yourself stop.
Route B: fine-tuned GPT-4o shown until April 2028
- If you already created the Azure resource while following Route A, keep it. Do not make duplicates without a reason.
- Otherwise, open Microsoft's Azure account page and create your subscription.
- Create a resource group and a resource that supports supervised fine-tuning for
gpt-4o-2024-08-06. - Check the current supported region, training type, quota and price before continuing.
Explain the Azure containers
Your subscription is the billing and access boundary. A resource group is a folder for related Azure items. A resource is the secure Azure service that holds model access, keys and jobs. A Foundry project is a workspace view used to build and manage AI work.
How do I check that this resource can fine-tune GPT-4o?
- Open your intended project in Microsoft Foundry.
- Open Build, then Fine-tune.
- Start creating a fine-tuning job, but do not submit it.
- Look for Supervised fine-tuning and the exact base model
gpt-4o-2024-08-06.
If that exact model and method are available to select, this check has passed. You can close the unfinished job and continue. If they are absent, stop and check Microsoft's current supported models, regions and training types before creating another resource.
Safe stopping point: creating the account and resource does not require you to submit a training job today.
- Download the starter pack below.
- Open READ-THIS-FIRST.txt.
- Read the training and validation examples yourself.
- Use them unchanged, edit them or replace them with conversations you have chosen and are entitled to use.
Low Spoons generic GPT-4o-style starter pack
Includes 31 training examples, 6 separate validation examples, a personalisation prompt, plain instructions and held-out test prompts.
You do not have to use all 31 examples. You can choose only the ones that fit, rewrite them, replace them with your own reviewed examples or remove complete examples to make the file smaller.
Choose, edit or reduce the examples
Each line of the training file is one complete example. Open the file in a plain-text or code editor and keep the lines you want. You may edit the words inside an example or delete a complete line. Do not delete half a line or leave broken brackets, quotation marks or commas.
If you add personal examples, you can replace generic ones instead of continually making the file longer. Keep the separate validation examples out of the training file, and keep the final test prompts out of both.
Check Microsoft's current minimum-example requirement for the model before reducing the file. A smaller file is cheaper, but variety still matters. Try to keep examples covering several qualities you value, such as warmth, humour, practical help, honest disagreement, repair and uncertainty.
What is inside this starter dataset?
The examples teach broad interaction qualities often valued in GPT-4o: warmth without clinical language, attention to incomplete thoughts, humour that fits the moment, practical collaboration, gentle disagreement, honest uncertainty, natural repair and respect for the user's agency.
They are original and intentionally generic. They contain no personal identity, memories or private conversation material. They are not GPT-4o's hidden training data and they cannot guarantee an exact recreation of GPT-4o.
Thank you and credit
This Low Spoons guide brings together the two Azure routes I had already tested and written about. Zyeine's GitHub guide, published afterwards, included one additional feature I particularly liked: a downloadable starter JSONL instead of an empty page.
That felt especially valuable here, because reducing executive load means removing avoidable setup wherever we can. With thanks to Zyeine for the download idea, I wrote a new, deliberately generic starter pack that readers can use unchanged or adapt. The Azure instructions and routes continue directly from my two earlier tested guides.
Safe stopping point: downloading and reviewing files costs nothing and uploads nothing to Azure.
- Open Fine-tuning in Microsoft Foundry and create a new supervised fine-tuning job.
- Choose base model
gpt-4o-2024-08-06. - Choose Developer training if it remains available and its displayed price fits your budget. It uses spare capacity, so it can queue.
- Upload the approved training JSONL and separate validation JSONL.
- For this first small job, use one epoch. Leave batch size and learning rate at their defaults.
- Switch automatic deployment off. Check this twice.
- Read the displayed price, then submit only when you are ready to spend that amount.
Why is automatic deployment off?
The completed custom model can remain stored without being continuously served. Automatic deployment can start an hourly hosting charge as soon as training finishes. Keeping it off lets you inspect the completed job before deliberately starting a live deployment.
Open Microsoft's current fine-tuning instructions
Microsoft's supervised fine-tuning guide documents the current schema, supported options, upload process and job controls. Use the interface and price displayed in your account if it differs from an older screenshot.
Safe stopping point: after submission, let the job queue. Do not resubmit merely because the time estimate changes.
- Wait until the job status says it has completed.
- Record the job ID, base snapshot, training type, epoch count, billed training tokens, custom-model identifier, date and time.
- Keep the training and validation files with that record so the experiment can be reproduced.
How long and how much did the tested job take?
My small Developer-tier job took about 45 minutes. Azure billed 2,000 training tokens and the calculated training cost at the displayed price was about 2.5 US cents. That is evidence from my job, not a price promise for yours.
Supervised fine-tuning cost depends on the training tokens, epoch count, training type, model and current rate. Waiting in a queue is not itself the training bill.
Safe stopping point: a completed custom model can remain stored while no deployment is running.
- Read the current hourly hosting price shown for the supported deployment type.
- Create a deployment at the smallest supported capacity.
- Open the custom-model details and record the retirement date shown on your model.
- If you are not evaluating it now, delete only the deployment.
- Return to the deployment list and prove it is empty.
Why does this guide say April 2028?
My completed Azure model displayed 13 April 2028 at 1:00 AM. Microsoft's generic lifecycle table showed a different earlier date for fine-tuned deployments. I report the date displayed on my actual model without promising that every account or future Microsoft policy will be identical.
Use my full tested April 2028 guide for the redacted evidence, date discrepancy and complete cost explanation.
Safe stopping point: once the deployment list is empty, the custom model remains stored but that deployment's hourly hosting meter is no longer running.
- Create a supported deployment only for a planned evaluation period.
- Use the held-out prompts from the starter pack. They were not included in training or validation.
- Give the same prompt and system instructions to the unchanged base model and the fine-tuned model.
- Compare warmth, judgement, humour, disagreement, repair, honesty, practical reasoning and memory boundaries.
- Delete the evaluation deployment when you finish, unless you knowingly accept the continuing hosting cost.
What should remain outside fine-tuning?
- System instructions can hold identity, values and boundaries.
- External memory can hold changing facts and shared history.
- The current conversation holds what is happening now.
- Fine-tuning is best used for durable response patterns demonstrated by reviewed examples.
Give the preparation to a coding AI
I want help preparing and validating an Azure GPT-4o supervised fine-tuning experiment. Do not submit, deploy, delete or spend anything until I explicitly approve that exact action. Requirements: - Use base snapshot gpt-4o-2024-08-06. - Check Microsoft's current schema, supported region, quota and price. - Convert only examples I have reviewed and approved. - Keep training, validation and final test examples separate. - Validate every JSONL line locally. - Remove secrets and flag sensitive third-party information. - Do not invent shared history or private quotations. - Keep automatic deployment off. - Record the job details and displayed costs. - If I later approve a temporary deployment, remind me that hourly hosting begins and help me verify its deletion afterwards.
Finished: you have a stored custom model, an evidence-based date from your own account and a test plan that does not grade the model on examples it already saw.
Official sources and fuller guides
- Ellivien's complete tested April 2027 Azure guide
- Ellivien's complete tested April 2028 fine-tuning guide
- Microsoft's model retirement schedule
- Microsoft's supervised fine-tuning guide
- Microsoft's fine-tuning cost guide
- Microsoft's live Azure OpenAI pricing
- Microsoft's Azure OpenAI data, privacy and security information
Checked 18 October 2026. Azure interfaces, prices, quota, regions and lifecycle dates can change. The dates in this guide report Microsoft's published base-model schedule and the retirement metadata displayed on my tested custom model.
Comments
Post a Comment