Azure fine-tuned GPT-4o retirement: 1 October 2027, GPT-4.1 on 14 October

Summary. Microsoft Foundry retires fine-tuned models in two steps. Deployments of fine-tuned gpt-4o (2024-08-06) and gpt-4o-mini (2024-07-18) stop on 1 October 2027, and fine-tuned gpt-4.1, gpt-4.1-mini and gpt-4.1-nano (2025-04-14) stop on 14 October 2027. Training can end six months earlier: no earlier than 1 April 2027 for the gpt-4o pair and 14 April 2027 for the gpt-4.1 family, and only for existing customers; for everyone else, training ends when the base models retire on 14 April 2027. Fine-tuned o4-mini shows 19 November 2027 on the schedule but 16 October 2027 on Microsoft's lifecycle page, so plan to the earlier date. Microsoft's guides describe no way to download the weights, idle fine-tuned deployments are deleted after 15 days without calls, and stored completions, the old portal route to distillation, were set to retire on 15 October 2026. The harder question is where to go next. Azure's fine-tuning table, updated 29 October 2026, lists no OpenAI model newer than gpt-5, which offers only invitation-only reinforcement fine-tuning on a base that retires on 9 February 2027. Each fine-tuned gpt-4.1 deployment costs $1.70 an hour to host, about $1,241 a month. Test a prompt-only replacement first, treat any GPT-4.1 retrain as a bridge, and move lasting custom-model work to open-weight bases.

About this article. This guide is for CTOs, ML engineers and platform owners running fine-tuned GPT-4o, GPT-4.1 or o4-mini models on Azure OpenAI in Microsoft Foundry. We checked every date, price and capability on 11 November 2026 against Microsoft Learn (the retirement schedule in its HTML and Markdown views, the lifecycle policy, and the fine-tuning overview, guides and cost pages), Microsoft's public retail prices API, Microsoft's GPT-6 launch post, and OpenAI's deprecations and model pages. Vendors change schedules and prices often, so re-check before you commit; no vendor paid for inclusion. Researched with AI assistance and reviewed by our editorial team.

The two clocks: training and deployment

Microsoft's lifecycle policy splits fine-tuned retirement into two phases. After training retirement, a model can no longer be fine-tuned, but models already trained stay deployable. At deployment retirement, inference and deployment calls return errors. Unless a date is stated, training retires no earlier than the base model, and the stated training dates apply to existing customers only; for anyone else, training ends at base-model retirement.

The table combines Microsoft's retirement schedule with OpenAI's deprecations page.

Three points shape the plan.

Azure is now the last home for these fine-tunes. OpenAI's API shuts down GPT-4o, GPT-4o mini, GPT-4.1 and GPT-4.1 mini on 14 January 2027, their fine-tuned versions go with them, and existing customers lose the ability to create new fine-tuning jobs there on 6 January 2027. Azure keeps serving its own fine-tunes for eight to nine months longer. That helps only models trained on Azure: an OpenAI fine-tune has to be retrained from your own data, which our guide to the end of OpenAI fine-tuning covers.

The retraining window closes in April. If your data has drifted and you want a refreshed GPT-4.1 fine-tune to carry you to October, train it before mid-April 2027. No earlier than means Microsoft may keep training open longer; it does not promise to. Eligibility is also narrower than it looks. Microsoft's lifecycle policy counts a subscription as an existing customer only if it has ever deployed that model version, all five base versions already sit in the Deprecated stage that closes them to other subscriptions, and a new subscription in the same tenant does not inherit access. Before planning any retrain, read the deprecation.fineTune field from the Models API for your region.

The o4-mini dates conflict. The schedule gives fine-tuned o4-mini training to 19 May 2027 and deployments to 19 November 2027; the lifecycle page, last updated 24 July 2026, gives 16 April and 16 October 2027. Each page read the same in its HTML and Markdown views on 11 November 2026, so the conflict is between the pages, not the formats. The base o4-mini model retires on 19 November 2026, eight days after our check. Plan to 16 October 2027 until Microsoft makes the pages agree.

Base-model dates and replacements are in our companion guide to the Azure GPT-4.1 and GPT-4o retirement on 14 April 2027.

What happens to your fine-tuned model and its weights

Five behaviours matter for planning.

No automatic upgrade. Microsoft's migration guide puts fine-tuned workloads out of scope for auto-upgrade: they carry their own training and deployment clocks, and adapting them means distillation or retraining rather than prompt and schema changes. When the deployment date passes, calls fail. Nothing moves your model to a newer base.

Weights stay inside Azure. Microsoft's classic fine-tuning guide says Azure fine-tunes with low-rank adaptation (LoRA). Its current supervised fine-tuning guide covers creating a job, inspecting checkpoints and downloading a results.csv file of training metrics; neither guide describes a way to download the trained weights of an OpenAI model. Plan as if the trained artefact disappears on the deployment date. What you keep is what you hold outside Azure: training and validation JSONL, grader code, evaluation sets and metrics.

Idle deployments are deleted early. A fine-tuned deployment that receives no chat completions or Responses API calls for a continuous 15-day period is deleted. The underlying model is not deleted and can be redeployed, in practice until deployment retirement. Each deployed fine-tune also incurs an hourly hosting charge whether or not it receives traffic.

Checkpoints can move, dates cannot. A finished job leaves the final model plus the two previous epochs as checkpoints, and a preview API copies a checkpoint to another region or subscription within the same tenant. That helps consolidate scattered fine-tunes before you retire them, but a copy keeps the same base model and the same October date.

Stored-completions datasets may be stuck. Microsoft set 15 October 2026 as the date the stored completions UI and API stop working. Existing stored completions cannot be exported or migrated, the data follows a rolling 30-day retention window, and jobs already built from those datasets keep working. Distillation training files built from stored completions cannot be accessed directly or downloaded. If your fine-tune came from portal distillation, you may hold no copy of its training data. Check now, and start capturing fresh examples through your own logging.

What you can still fine-tune on Azure

Microsoft's fine-tuning overview, updated 29 October 2026, lists every supported combination; the retirement column comes from the schedule.

Read from the top, the problem is plain. No GA OpenAI model newer than GPT-4.1 supports supervised fine-tuning on Azure. The newest tunable OpenAI model, gpt-5, takes RFT only, by invitation, on a base that leaves the schedule on 9 February 2027, as our Azure GPT-5 retirement guide explains. GPT-6 Astra, Sol and Luna, and the newer GPT-6.1 Sol and Terra, are in Azure's model catalogue but not in its fine-tuning table, and OpenAI's model pages for GPT-6 Luna and GPT-6 Sol list fine-tuning as not supported. Two of the GA open models, Llama-3.3-70B-Instruct and Ministral-3B, already carry near retirement dates.

The open-weight rows are where custom models still have a future on Azure. gpt-oss-20b and Qwen3-32B are GA for SFT with Global Standard serverless deployment and no date on the schedule yet. The preview rows add RFT and Programmed Training, where you write the training loop in Python and Foundry runs it; Programmed Training needs access approval, and Microsoft does not recommend preview features for production. Fireworks on Foundry can also import model files you trained elsewhere, which keeps an open-weight adapter portable in a way an Azure OpenAI fine-tune is not.

Four migration paths

Path 1 should be the default test. Many fine-tunes from 2024 and 2025 fixed problems newer models handle without training: strict JSON, a house style, classification into a fixed set. GPT-6 Luna on Azure supports structured outputs and a 1,050,000-token context window, which leaves room for many worked examples. Score a prompt-only version on your frozen evaluation set before spending anything on training. Set reasoning.effort explicitly: OpenAI's model page says Luna defaults to medium, and Chat Completions supports function calling only when effort is none.

Path 2 is a bridge, not a destination. A gpt-4.1-mini retrain costs $5 per million training tokens with Global training, or $2.50 in the Developer tier. Developer training uses idle capacity, can be pre-empted and carries no data-residency guarantee. Whatever you train still stops on 14 October 2027.

Path 3 is Microsoft's own suggestion for fine-tuned workloads. The overview lists distillation through teacher-generated SFT data under Managed Training, and the migration guide names distillation or retraining as the routes for fine-tunes. The pattern: run a strong teacher (your current fine-tune, or GPT-6 Sol) over representative inputs, filter the outputs with graders, then train a smaller open-weight student. With stored completions gone, Microsoft points to Agent Traces for building fine-tuning datasets from agent runs; traces sit in your Application Insights resource, and the portal searches the last 90 days. For non-agent workloads, log teacher inputs and outputs yourself, and check your agreement's terms on training with model outputs before you start.

Path 4 is covered elsewhere. Google, AWS, Together, Fireworks and self-hosted LoRA options are priced in the OpenAI fine-tuning alternatives guide linked above.

Worked cost example

Illustrative monthly arithmetic using Azure Global prices from Microsoft's retail prices API (North Central US meters) and GPT-6 short-context Global Standard prices from Microsoft's launch post. eCorpIT did not benchmark these models; the prompt growth is our assumption, so measure your own.

A team runs four fine-tuned deployments: three gpt-4.1-mini fine-tunes that together handle 600M input and 60M output tokens a month, and one gpt-4.1 fine-tune handling 400M input and 40M output.

Three observations. Hosting is 77% of today's bill ($4,964 of $6,420), so idle or duplicate fine-tunes are the first thing to cut, whatever path you choose. Path 1 roughly halves the bill even with prompts three times longer, before any caching: if two-thirds of the Sol prompt were a repeated prefix billed at the $0.20 cached-input rate, the gpt-4.1 line would fall from $2,800 to $1,360 (800M x $0.20 + 400M x $2.00 + $400 of output), before cache-write charges of $2.50 per million. Reasoning tokens are billed as output tokens, which pushes the other way, so measure them at the effort level you deploy.

For Path 2, a refresh before April with a 10M-token dataset over three epochs (30M training tokens) costs $150 per gpt-4.1-mini model and $750 for gpt-4.1 at Global rates, $1,200 for all four, or $600 in the Developer tier; Microsoft bills SFT as training tokens times epochs times the per-token rate. Hosting continues at $4,964 a month until October.

For Path 3, generating 20,000 teacher examples at 2,000 input and 500 output tokens each on GPT-6 Sol costs 40M x $2 + 10M x $10 = $180. Azure's retail feed lists fine-tuning meters for both open students. In North Central US, excluding Data Zone meters, gpt-oss-20b costs $3.60 per million training tokens, $0.30 an hour to host, and $0.07 input and $0.30 output per million tokens; Qwen3-32B costs $3.20 to train, $0.30 an hour to host, and $0.30 input and $1.20 output. Training a gpt-oss-20b student on those 20,000 examples (50M tokens) over three epochs costs 150M x $3.60 = $540. If three gpt-oss-20b students replaced the three gpt-4.1-mini fine-tunes, hosting would be 3 x 730 h x $0.30 = $657 and inference 600M x $0.07 + 60M x $0.30 = $60, or $717 a month against $4,059 for those three today. Whether a 20B student matches your fine-tune is what the evaluation set must show.

A timeline for CTOs

Four habits make this smoother. Keep a frozen evaluation set per fine-tune, so the old model can be scored against its replacement on identical inputs. Shadow the new model on production inputs before any customer sees it, which Microsoft's migration guide recommends for regulated flows that cannot canary. Route traffic through a gateway so the cut-over is a configuration change; our AI gateway comparison covers the options. And delete each old deployment the day its traffic reaches zero, because hosting is billed by the hour.

India, EU and US notes

India. No fine-tuned GPT-4o or GPT-4.1 option on Azure keeps processing inside India. Standard (regional) fine-tuned deployments are offered only in East US 2, North Central US and Sweden Central, and the Data Zone option for fine-tuning is US only. Global Managed Training can be started from a South India project, but Microsoft says data and weights are copied to the region where training occurs, and Global Standard deployments may store custom weights temporarily outside the resource's geography. Teams whose DPDP Act assessment needs in-country processing should look at open-weight models they can host in an Indian region themselves.

EU. Sweden Central supports Standard deployments of the fine-tuned gpt-4o and gpt-4.1 families, plus Provisioned Throughput for fine-tuned gpt-4o-mini, gpt-4o and gpt-4.1. There is no EU Data Zone for fine-tuning. EU teams moving to a prompt-only GPT-6 model can use the EU Data Zone, where Luna costs $0.12 input and $0.60 output per million short-context tokens.

US. North Central US and East US 2 cover Standard deployments, with model support varying by region, and the US Data Zone covers both training and deployment for the OpenAI models in the table. Microsoft's lifecycle policy notes that not every commercial model reaches the Azure Government clouds, so confirm fine-tuning support there with Microsoft.

How eCorpIT can help

eCorpIT is a Gurugram-based technology organisation, founded in 2021, assessed at CMMI Level 5 and MSME certified, with senior-led engineering teams working across AWS, Microsoft and Google platforms. Our AI evaluation and observability service builds golden-set regression evals that score prompt-only, retrained and distilled replacements against your current fine-tunes before you switch. Talk to us at /contact-us/.

Last updated: 11 November 2026.

References

1. Model retirement schedule, Microsoft Foundry, updated 23 September 2026 — Microsoft Learn

1. Foundry Models lifecycle and support policy, updated 24 July 2026 — Microsoft Learn

1. Deploy a fine-tuned model (classic), updated 8 October 2026 — Microsoft Learn

1. Migrate from stored completions to Responses API and Agent Traces (classic) — Microsoft Learn

1. Fine-tuning in Microsoft Foundry (overview and supported models), updated 29 October 2026 — Microsoft Learn

1. Azure retail prices API, Azure OpenAI meters in North Central US — Microsoft

1. Deprecations — OpenAI

1. Model retirement schedule, Markdown view — Microsoft Learn

1. Model migration process in Microsoft Foundry — Microsoft Learn

1. Customize a model with fine-tuning (classic) — Microsoft Learn

1. Fine-tune a model with supervised fine-tuning in Microsoft Foundry — Microsoft Learn

1. Stored completions and distillation (classic) — Microsoft Learn

1. Foundry Models sold directly by Azure, updated 29 October 2026 — Microsoft Learn

1. GPT-6 Luna model page — OpenAI

1. GPT-6 Sol model page — OpenAI

1. Deploy fine-tuned models in Microsoft Foundry — Microsoft Learn

1. GPT-6 Astra, Sol, and Luna for production agents in Microsoft Foundry, 22 September 2026 — Microsoft Azure Blog

1. Azure OpenAI reasoning models — Microsoft Learn

1. Manage fine-tuning costs in Microsoft Foundry — Microsoft Learn

1. Azure retail prices API, Azure OpenAI OSS Models meters in North Central US — Microsoft

1. Azure retail prices API, FT Models meters in North Central US — Microsoft