OpenAI fine-tuning deprecated: new jobs end 6 January 2027, where to go next

Summary. OpenAI is winding down self-serve fine-tuning. Active existing customers can no longer create jobs from 6 January 2027, after the platform closed to new organisations on 7 May 2026 and to inactive ones on 2 July 2026. Existing fine-tunes serve only until their base model goes, and gpt-4.1 and gpt-4.1-mini shut down with GPT-4o and GPT-4o mini on 14 January 2027, eight days later. gpt-4.1-nano and fine-tuned o4-mini already went on 23 October 2026, and the named replacement, GPT-6 Luna, cannot be fine-tuned. The realistic options: retrain GPT-4.1 on Azure, if your subscription has already deployed that version, at $5 per million training tokens for gpt-4.1-mini plus $1.70 an hour hosting, with deployments ending on 14 October 2027; tune Gemini 3.5 Flash-Lite at $3 per million with base-model inference prices; use Amazon Bedrock's Nova 2 Lite at $3.78 per million or reinforcement fine-tuning on gpt-oss-20b; or train LoRA adapters at Together ($0.40 per million for gpt-oss-20B) or Fireworks ($0.50 per million up to 16B) and serve them on dedicated GPUs or your own vLLM. First test whether a prompt on a current model already beats your fine-tune.

About this article. This guide is for engineering and ML leads running fine-tuned OpenAI models. Dates, prices and features were checked on 1 November 2026 against OpenAI's deprecations, guides and pricing, Microsoft Learn and the Azure retail prices API, Google Cloud and AWS Bedrock documentation and prices, and Together AI, Fireworks and vLLM documentation. Vendors change prices often, so re-check before you commit; this is not procurement or legal advice, and no vendor paid for inclusion. Researched with AI assistance and reviewed by our editorial team.

What OpenAI is ending, and why 14 January matters more

OpenAI notified developers on 7 May 2026. With its model shutdowns, the timeline from its deprecations page is:

OpenAI's supervised fine-tuning guide confirms both dates. A job you start in the first week of January produces a model that lives for days.

Two details catch teams out. OpenAI's pricing page still shows reinforcement fine-tuning (RFT) on o4-mini-2025-04-16 at $100 per training hour, but that model shut down on 23 October 2026 and the RFT guide names no other base, so do not budget from that row. And the replacement OpenAI names for the GPT-4 series, gpt-6-luna, lists no fine-tuning endpoint, as our GPT-4o and GPT-4.1 shutdown guide explains. After 14 January, OpenAI's self-serve platform offers no route to a custom model.

What happens to your existing fine-tuned models

OpenAI dates follow its deprecations page and guides. Azure dates follow Microsoft's retirement schedule, last updated on 23 September 2026.

Azure's later training dates apply to existing customers only; for anyone else, training ends when the base model retires. Microsoft's lifecycle page, last updated on 24 July 2026, still shows earlier o4-mini dates of 16 April and 16 October 2027, so confirm that row before relying on it.

OpenAI's fine-tuning guides describe no way to download a fine-tuned model's weights. What you own is the training data, the hyperparameters and the evaluation set. Export all three now. If your fine-tunes used RFT graders, copy those definitions before the Evals shutdown on 30 November 2026 as a precaution.

First, check whether you still need fine-tuning

Many older fine-tunes exist to make a small model follow a format or label set. Current models with structured outputs and cheap cached prefixes may match them without training. As a baseline, GPT-6 Luna costs $0.10 per million input tokens and $0.50 per million output tokens; it defaults to medium reasoning effort, so set effort none for a fair cost test.

Run your fine-tuned model and two or three prompt-only candidates against the same golden set before deciding. Fine-tuning still earns its cost when:

• latency targets need a small model that cannot follow long instructions reliably;

• volume is high enough that removing 1,000 tokens of examples from every prompt matters;

• outputs depend on domain vocabulary or behaviour that is hard to describe in a prompt;

• you need weights you control, for data residency or to escape vendor retirement schedules.

Option 1: Azure OpenAI in Microsoft Foundry

Azure is the one platform in this comparison that can still train the same OpenAI models with the same chat-format JSONL. Microsoft's fine-tuning overview, last updated on 29 October 2026, lists supervised fine-tuning (SFT) for gpt-4o-mini, SFT and direct preference optimisation (DPO) for gpt-4o 2024-08-06 and the gpt-4.1 family, RFT for o4-mini, and invitation-only RFT for gpt-5. It also lists SFT for four open models on Global Standard (Llama-3.3-70B-Instruct, Qwen3-32B, gpt-oss-20b and Ministral-3B), plus preview training for newer open models served on managed compute or Fireworks on Foundry.

Prices below are per million tokens from Microsoft's public retail prices API for North Central US.

Regional training costs about 10% more. The hosting fee changes the maths: Microsoft adds an hourly fee for as long as a Standard or Global Standard fine-tune stays deployed, which is $1,241 a month per deployment at 730 hours. Developer deployments carry no hourly fee, but they are for evaluation only, disappear after 24 hours and are not offered for the gpt-4o family.

Two caveats. First, Azure lists these base versions as Deprecated, and its lifecycle policy closes Deprecated models to any subscription that has never deployed that version; a new subscription in the same tenant does not inherit access. Confirm that your subscription qualifies before you plan a retrain. Second, Azure buys time, not a destination: fine-tuned GPT-4.1 deployments end on 14 October 2027, and base gpt-5, the gated RFT option, retires on Azure on 9 February 2027, as our Azure GPT-5 retirement checklist covers.

Option 2: Google Gemini Enterprise Agent Platform

Google's enterprise platform, formerly Vertex AI, supports SFT for Gemini 3.5 Flash-Lite, 3.5 Flash and 3.1 Flash-Lite. Its guide also still lists Gemini 2.5 models, whose retirement date of 20 October 2026 has passed. Gemini 3.5 Flash also supports reinforcement learning fine-tuning. The table follows Google's pricing and model versions pages.

Open models tune at list prices such as $6.72 for Llama 3.3 70B and $6.57 for Qwen 3 32B. Training tokens are dataset tokens times epochs. From Gemini 3 onwards, a tuned endpoint costs the base price in the global region and 1.5 times the base price elsewhere. Google's pricing page lists no hourly hosting fee for tuned Gemini endpoints.

Google's guide says to set the thinking level to MINIMAL for Gemini 3 tuned tasks and to skip controlled generation (response schemas) at inference because it can lower quality. Supervised fine-tuning is not a Covered Service under Google's service level agreements. Google says listed retirement dates can move later but not earlier; plan on a tuned model living no longer than its base.

Option 3: Amazon Bedrock

Bedrock's fine-tuning table looks long, but most rows are closed in practice. It lists Claude 3 Haiku, which reached end-of-life in us-west-2 on 10 September 2026, Llama 3.2 models that are past end-of-life, and Llama 3.1 and 3.3 models that are now Legacy with end-of-life on 30 March 2027. Bedrock's lifecycle rules block new fine-tuning jobs on Legacy models. That leaves two routes:

• Amazon Nova. Nova 2 Lite, Nova Micro, Nova Lite and Nova Pro support fine-tuning in us-east-1. AWS lists Nova 2 Lite SFT at $0.00378 per 1,000 tokens ($3.78 per million), RFT at $80 per hour and storage at $1.95 per custom model per month, and it prices on-demand inference for custom Nova models the same as base Nova.

• Open weights with an OpenAI-shaped API. Bedrock's RFT covers Nova 2 Lite, gpt-oss-20B and Qwen3 32B. For the two open models, AWS exposes OpenAI-compatible fine-tuning endpoints on bedrock-mantle.us-west-2.api.aws, with inference through the Responses or Chat Completions APIs and no separate deployment step. gpt-oss-20b RFT costs $80 per training hour, then $0.09 per million input and $0.39 per million output tokens, plus $1.95 a month storage.

AWS documents the OpenAI-compatible route for RFT with Lambda reward functions, so check whether it fits a supervised JSONL workflow before you plan around it.

Option 4: Open-weight models on Together AI or Fireworks

Both train LoRA adapters on open-weight models and let you download the result, which removes vendor retirement from the equation.

Together AI prices LoRA SFT per million training tokens: $0.40 for gpt-oss-20B, $2.50 for gpt-oss-120B, $2.03 for Llama 3.3 70B and $1.05 for Qwen3.5 27B or Gemma 4 31B, with minimum charges from $4 per job. Validation tokens count too, and DPO costs 2.5 times the SFT rate. Serving has changed: Together's documentation says serverless LoRA inference has been discontinued, so adapters run on dedicated endpoints, from $5.49 per H100 hour. Jobs output an adapter plus merged weights you can download.

Fireworks charges for LoRA SFT by model size: $0.50 per million tokens up to 16B parameters, $3.00 for 16.1B to 80B, $6.00 for 80B to 300B and $10.00 above 300B. LoRA DPO and full-parameter SFT each cost twice the LoRA SFT rate. Its deployment documentation limits trained LoRA models to on-demand (dedicated) deployments, at $8.00 per H100 hour, with a 1.5x premium for region-restricted deployments. Live merge matches base-model speed for one adapter; multi-LoRA shares one deployment across many adapters, and Fireworks says time to first token is often 10% to 30% higher.

Option 5: Self-host LoRA adapters with vLLM

To own the serving stack as well, train an adapter on Together or Fireworks and serve it yourself. vLLM loads adapters with --enable-lora and --lora-modules, exposes each one as a model name on an OpenAI-compatible server, and can hold several adapters on one base model.

            vllm serve meta-llama/Llama-3.1-8B-Instruct \
  --enable-lora \
  --lora-modules ticket-classifier=/models/adapters/ticket-classifier
# Call it like any OpenAI model: "model": "ticket-classifier"
          

GPU time is the main cost. Together's GPU clusters list an on-demand H100 at $3.99 an hour ($2,913 for a 730-hour month) and a preemptible H100 at $1.99. Our comparison of Modal, Baseten, RunPod and Replicate covers managed GPU options.

Comparison table

Prices and limits are cited above. Not documented means the fine-tuning documentation describes no weight export.

Decision table

Worked cost example

This is illustrative arithmetic from published prices; eCorpIT did not benchmark these models. A support-ticket classifier: 10,000 training examples of 1,500 tokens for 3 epochs (45 million training tokens), then 1 million requests a month at 1,000 input and 100 output tokens (1,000 million input, 100 million output). The prompt-only option adds 1,000 tokens to every request. OpenAI's fine-tune rates come from the expanded table on its pricing page, and the Nova row assumes Bedrock's in-region $0.33 and $2.75 per million.

What the numbers say:

• Azure's hosting fee dominates at this volume. For gpt-4.1-mini, $1,241 of the $1,801 is the deployment fee, so consolidate fine-tunes or shut idle ones down.

• Per-token endpoints win at moderate volume. Tuned Gemini 3.5 Flash-Lite works out at $0.50 per million blended tokens. A $4,008 dedicated H100 only breaks even near 8,000 million tokens a month, about seven times this workload, assuming one GPU handles it.

• Bedrock's open-weight RFT is the cheapest fine-tuned option per token at about $0.12 per million blended, but RFT needs a reward function, and quality on your task decides.

• The prompt-only baseline is hard to beat on cost. Only Bedrock's RFT option undercuts it here, so fine-tuning has to win on quality, latency or control.

• Model size moves the Fireworks line. A 20B model costs 45 x $3 = $135 to train.

India, EU and US data notes

India. No managed option here offers training with an India residency guarantee. A South India Foundry project can start Azure Global training, but data and weights are copied to wherever training runs, and Standard deployments are listed only in East US 2, North Central US and Sweden Central. Bedrock customisation runs in us-east-1 and us-west-2. Google stores tuning data in the job's region but says computation could be offloaded to other US or EU regions. Teams that want training data and inference to stay in India should self-host adapters in an Indian cloud region; our DPDP engineering playbook covers retention and processing records.

EU and US. Azure lists Sweden Central as a Standard deployment region for the GPT-4o and GPT-4.1 families, and its data zone training is US only. On Google, confirm that an EU tuning job region exists for your model and budget 1.5 times base price for non-global tuned Gemini 3 endpoints. Fireworks charges 1.5 times for region-restricted deployments.

Migration checklist

1. Inventory every fine-tuned model ID (the ft: prefix) in code, gateway configs and environment variables, with its owner.

1. Export before 6 January 2027: training and validation JSONL, hyperparameters and result metrics. Copy RFT grader definitions before 30 November 2026.

1. Capture reference outputs. Log a few thousand production inputs and outputs from each fine-tune before 14 January for your golden set.

1. Test prompt-only first on GPT-6 Luna and Gemini 3.5 Flash-Lite against that golden set.

1. Retrain on one or two targets and compare task accuracy, schema validity, latency and cost per successful task.

1. Route through a gateway such as those in our AI gateway comparison, shadow traffic, then canary before 14 January.

1. Calendar the new base model's retirement date on day one; our AI model retirement policy comparison lists each platform's rules.

1. Keep the recipe portable. Store datasets, configs and evals in your own repository.

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 open-model fine-tuning service exports your OpenAI training assets, builds golden sets from production traffic, tests prompt-only baselines and retrains on the target that fits your cloud, data-residency and cost limits. Talk to us at /contact-us/.

Last updated: 1 November 2026.

References

1. Deprecations — OpenAI API

1. Supervised fine-tuning — OpenAI API

1. GPT-6 Luna model — OpenAI API

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

1. Model retirement schedule, Microsoft Foundry — Microsoft Learn

1. Foundry Models lifecycle and support policy — Microsoft Learn

1. Gemini Enterprise Agent Platform generative AI pricing — Google Cloud

1. Amazon Bedrock pricing — AWS

1. Pricing — Together AI

1. Pricing — Fireworks AI

1. Pricing — OpenAI API

1. Reinforcement fine-tuning — OpenAI API

1. Vision fine-tuning — OpenAI API

1. Fine-tuning in Microsoft Foundry, updated 29 October 2026 — Microsoft Learn

1. Fine-tuning cost management, Microsoft Foundry — Microsoft Learn

1. About supervised fine-tuning for Gemini models — Google Cloud

1. Model versions and lifecycle — Google Cloud

1. Customize a model with fine-tuning in Amazon Bedrock — AWS

1. Model lifecycle, Legacy and end-of-life models — AWS

1. Customize a model with reinforcement fine-tuning in Amazon Bedrock — AWS

1. Fine-tune open-weight models using OpenAI-compatible APIs — AWS

1. LoRA vs. full fine-tuning — Together AI Docs

1. Deploying trained models — Fireworks AI Docs

1. LoRA adapters — vLLM Docs