OpenAI Is Winding Down Fine-Tuning: Where to Move Your Fine-Tunes (October 2026)

Last updated: October 18, 2026

Short answer: OpenAI's self-serve fine-tuning is closing in stages. It already shut out organizations that never fine-tuned (May 7, 2026) and organizations with no fine-tuned inference in the previous 60 days (July 2, 2026). On January 6, 2027, "Active existing customers will no longer be able to create new fine-tuning jobs." Your existing fine-tunes keep serving until their base model shuts down, and every base model OpenAI's guides and pricing page still list as tunable has a shutdown date on or before January 14, 2027. Two of those tunable bases go much sooner, on October 23, 2026: o4-mini (the only reinforcement fine-tuning base) and GPT-4.1 nano.

So the real question is not whether to move, but where. Fine-tuning still exists on Microsoft Foundry, Vertex AI and Amazon Bedrock for closed and open models, and on Together AI, Fireworks AI, Nebius and Alibaba Cloud for open-weight models. Only some of them let you download the result, and that turns out to be the property that matters most.

Everything below comes from our fine-tuning directory: 12 providers, each value stored with a verbatim quote from the provider's own docs or pricing page, read on October 18, 2026. OpenAI's dates also come from our deprecations tracker. Prices stay in each provider's own unit. Where a provider's page does not say something, we write "not stated".

What changed: the dates#

OpenAI's deprecations page carries the schedule. Today is October 18, 2026.

The rule that ties those rows together is in the same section: "Inference on fine-tuned models will be disabled only when the underlying base model is deprecated." And for the January 14 group: "Fine-tuned versions of these models follow the same shutdown schedule as their underlying base models."

Now line that up with what is still tunable. The supervised fine-tuning guide lists gpt-4.1-2025-04-14, gpt-4.1-mini-2025-04-14 and gpt-4.1-nano-2025-04-14. The pricing page adds gpt-4o-2024-08-06 and gpt-4o-mini-2024-07-18. The reinforcement fine-tuning guide says it is available "only for o4-mini". GPT-4.1 nano and o4-mini go on October 23. Everything else goes on January 14, 2027.

That means a job you start on January 5, the last day before the cutoff, produces a model with at most nine days to live. Every OpenAI fine-tuning guide we read carries the same line: "OpenAI is winding down the fine-tuning platform."

The replacements OpenAI recommends are base models: gpt-6-luna for the January 14 group, and gpt-5.6-terra, gpt-5.6-sol or gpt-5.6-luna for the October 23 removals. None of them appears in the fine-tuning guides' lists of tunable models. On OpenAI, the path forward is a newer base model plus prompting, not a new fine-tune. Our shutdown calendar for 2026-2027 puts these dates next to every other provider's.

What still runs on OpenAI, and for how long#

If you have fine-tunes in production, sort them by base model first. Our OpenAI fine-tuning page has the full offering with quotes.

Keep serving until the base shuts down if your fine-tune is doing its job. You have 88 days, which is enough to build an eval set, test two or three destinations and cut over behind a flag. The case for moving now is any plan that depends on retraining. New training on OpenAI costs $25.00 per 1M training tokens on GPT-4.1, $5.00 on GPT-4.1 mini, $3.00 on GPT-4o mini, and $100.00 per hour for o4-mini RFT, per the pricing page. Any of those models stops serving on January 14 at the latest.

Where fine-tuning still exists#

Hosted closed models#

Microsoft Foundry fine-tunes the same OpenAI models, and it gives them a longer tail. Its supported-models table lists managed SFT and DPO on gpt-4o (2024-08-06) and the GPT-4.1 family, SFT on gpt-4o-mini, and RFT on o4-mini. GPT-5 RFT "requires an invitation." Per Microsoft's model retirement schedule, as recorded in our deprecations tracker:

That is roughly nine more months than OpenAI gives the same models. Microsoft's two pages disagree on fine-tuned o4-mini, and our Foundry retirements post covers both dates and the separate training and deployment phases. Whether Foundry still accepts new training jobs on a base it lists as Deprecated is not stated in our data. Nothing we read describes moving an existing OpenAI fine-tune to Foundry either, so plan to retrain from your own data.

Foundry's pricing has two parts, training and hourly hosting. From the Azure OpenAI pricing page: GPT-4.1 training is $25/1M tokens on Global and $12.50/1M on Developer, with hosting at $1.7/hour on Global. o4-mini RFT training is $100/hour on Global and $50/hour on Developer. The Developer tier has no hourly hosting fee, but Microsoft describes its deployments as "For evaluation only; lasts 24 hours and has no availability SLA." Global training has a residency catch: "Data and weights are copied to the region where training occurs."

Vertex AI is where Gemini tuning lives now. The Gemini API is out: "we no longer have a model available which supports fine-tuning in the Gemini API or AI Studio," per Google's model tuning page. Vertex's supervised tuning page lists Gemini 3.5 Flash, Gemini 3.1 Flash-Lite, Gemini 2.5 Pro, Gemini 2.5 Flash-Lite and Gemini 2.5 Flash. Check the dates before you pick one, though. Our deprecations tracker shows Vertex retiring gemini-2.5-pro, gemini-2.5-flash and gemini-2.5-flash-lite on October 20, 2026, two days from now. Our Vertex shutdowns post covers what that means for tuned 2.5 models. That leaves Gemini 3.5 Flash and 3.1 Flash-Lite. The tuning page's anchors for those two are named as public preview, so we infer they are Preview, but the page text does not say.

Vertex prices training per "count" on its pricing page: $0.01 / 1,000 count for Gemini 3.5 Flash (which also supports reinforcement learning fine-tuning) and $0.003 / 1,000 count for Gemini 3.1 Flash Lite. Serving costs more than the base model: "For model inference starting from Gemini 3, tuned model endpoint prediction price will be 1.5 times of the base model." Tuned models serve only from the us and eu multi-region endpoints. Vertex also tunes open models: Gemma 3 27B IT at $6.83 / 1,000,000 count, Llama 3.3 70B at $6.72 and Qwen 3 32B at $6.57.

Amazon Bedrock offers supervised fine-tuning, reinforcement fine-tuning and distillation, per its custom models page. The SFT table covers Amazon Nova (Nova 2 Lite, Micro, Lite, Pro and Canvas), Llama 3.1, 3.2 and 3.3, and one Claude model. Llama 3.1 8B trains at $1.49 per 1M tokens, plus $1.95 per custom model per month for storage, and serves on no-commit Provisioned Throughput at $24.00 per model unit per hour (Bedrock pricing). Bedrock's pricing page shows Nova text-model training prices per Region, and we have not recorded them here, so check your Region before budgeting.

The simpler serving story is reinforcement fine-tuning on Nova 2 Lite, openai.gpt-oss-20b and qwen.qwen3-32b. Training is $80.00 per training hour for gpt-oss-20b and Qwen3 32B, and "customers can immediately use the resulting fine tuned model for on-demand inference": $0.09 input / $0.39 output per 1M tokens for gpt-oss-20b and $0.20 / $0.78 for Qwen3 32B, plus $1.95 a month storage.

Claude is the gap. The Claude API pricing page lists no fine-tuning. Anthropic's own cookbook walks through "finetuning Claude 3 Haiku on Amazon Bedrock", hosted with Provisioned Throughput, and Bedrock's table still lists anthropic.claude-3-haiku-20240307-v1:0:200k in us-west-2. But AWS's model lifecycle page, as recorded in our tracker, shows Claude 3 Haiku reaching end of life in us-west-2 on September 10, 2026, with its remaining Regions following on November 13, 2026. Bedrock also says "Distillation is not currently available for Anthropic models on Amazon Bedrock." If your OpenAI fine-tune was going to become a Claude fine-tune, the docs we read offer no current route.

Open-weight models: LoRA and full fine-tuning#

These providers train open-weight models, and that matters for the next section.

Together AI does LoRA, full-parameter and DPO fine-tuning. Prices per 1M training tokens from its pricing page, each job with a minimum charge: Llama 3.1 8B LoRA $0.34 (minimum $4.00), Llama 3.3 70B LoRA $2.03 or full $2.24, gpt-oss-120B LoRA $2.50 (minimum $6.00), DPO on Llama 3.1 8B $0.84. The page's two tables are unlabelled tabs, so which one is LoRA and which is full is inferred from tab order. Serving is on a dedicated endpoint, and "Some models can be fine-tuned but cannot be deployed as dedicated endpoints."

Fireworks AI prices managed SFT and DPO by base-model size, per 1M training tokens (pricing): LoRA SFT $0.50 up to 16B, $3.00 for 16.1B to 80B, $6.00 for 80B to 300B, $10.00 above 300B. Full-parameter SFT is double the LoRA rate in each tier. Its pricing page says "Serve fine-tuned models for the same price as base models", while its fine-tuning docs say an on-demand (dedicated) deployment "is the only supported method for serving trained models." Read both before you budget. A Serverless Training API with RL loops "is now generally available."

Nebius Token Factory runs "long-context supervised fine-tuning on full model weights or LoRA adapters" on gpt-oss, Qwen3, DeepSeek V4 Flash and Gemma 4 (overview). Deployment is dedicated endpoints only. Training prices are not stated in the pages we could read; its prices page sits behind a console login. Reinforcement fine-tuning is a "limited professional service".

Alibaba Cloud Model Studio fine-tunes Qwen in its Singapore region: $0.0075/Thousand Token for Qwen3.8-27B and $0.003 for Qwen3.5-9B (training overview). A fine-tuned model "can only be used after deployment", and for SFT the docs say the dataset "requires at least 1,000 high-quality fine-tuning data entries".

What is gone#

• Cohere: "Cohere's fine-tuning feature was deprecated on September 15, 2025" (docs).
• Mistral: the fine-tuning docs moved under deprecated features: "This feature is deprecated and is no longer actively supported." No end date is given (docs).
• Hugging Face AutoTrain: "This project is no longer maintained." It still runs as a Space billed per minute of hardware (docs).
• Gemini API and AI Studio: no tunable model since Gemini 1.5 Flash-001 was deprecated in May 2025.
• xAI and DeepSeek: we found no fine-tuning page in either API's docs. That is not proof neither offers it, only that we could not find one.

Weight export is the hedge#

Look at what just happened on OpenAI. A hosted fine-tune lives exactly as long as its base model, and the provider sets that date. A checkpoint you can download is different: you can serve it on another host, or on your own hardware, after the provider moves on.

Two caveats. "Not stated" is not "no", so check your provider's current docs before you rule one out. And a downloaded checkpoint is only as portable as its base model's license. Before you commit to an open-weight base, check its terms in our open-weight license table.

The practical rule: if a fine-tune is core to your product, train it on an open-weight base somewhere that lets you download the result. Even if you keep serving it on that provider, a copy of the weights means the next shutdown notice is a hosting decision, not a retraining project.

Migration checklist#

1. Inventory. List every fine-tuned model ID your code calls and map each to its base model's shutdown date. The OpenAI deprecations data has them. Anything on o4-mini or GPT-4.1 nano needs a fallback before October 23.
2. Secure your data. Gather every training and validation file you used, plus any you uploaded and no longer keep locally. Do not assume they will be retrievable after shutdown.
3. Build an eval set first. Pull real inputs from production, write down what a correct output looks like, and freeze the set. Every later decision is a score on this set.
4. Baseline the replacement base model. Run OpenAI's recommended replacement, or the newest base model at your target provider, with your best prompt and a handful of examples from your training data. If it matches your fine-tune, stop. You no longer need one.
5. Pick a destination by need. Same OpenAI models for longer: Foundry, with fine-tuned deployments retiring October 2027. Gemini: Vertex, on 3.5 Flash or 3.1 Flash-Lite. Open-weight with downloadable weights: Together or Nebius. Open-weight with on-demand serving straight after RFT: Bedrock.
6. Check data limits before uploading. Some examples from our data: OpenAI's DPO guide says "we only train on one-turn conversations"; Vertex allows "1GB for JSONL" and 131,072 input and output tokens per training example; Fireworks needs at least 3 examples; Alibaba asks for at least 1,000 SFT entries. We did not compare dataset schemas across providers, so validate a small file against each provider's format before converting everything.
7. Price serving, not just training. Foundry charges hourly hosting ($1.7/hour for GPT-4.1 on Global). Bedrock SFT models sit on Provisioned Throughput ($24.00 per model unit per hour for Llama 3.1, no commitment). Together, Fireworks, Nebius and Alibaba serve on dedicated deployments. Vertex charges 1.5 times the base price for tuned Gemini 3 models.
8. Check where your data goes. Foundry Global training copies data and weights to the training region. Bedrock says data "never leaves AWS's secure, governed environment during the customization process" for RFT. Our data retention table covers each provider's API terms.
9. Cut over behind a flag and keep the old fine-tune serving as a fallback until it shuts down.

Try these before you fine-tune again#

Many fine-tunes predate today's base models and today's prompt caching discounts, so test the cheaper options against your eval set before paying for training again:

• A long prompt with cached examples. Put your instructions and a few dozen worked examples in a fixed prefix and let prompt caching discount it. Our cost levers table has every provider's caching and batch rates with quotes, and the guide to cutting your API bill works through the math.
• Distillation. Use a strong model to generate training data for a small one. Bedrock runs this as a managed service: teacher calls are "billed at the on-demand inference rates of the teacher model", and synthesis can grow the dataset "to a maximum of 15k prompt-response pairs".
• RL on a small open model. If your task has a checkable answer, reinforcement fine-tuning a small open-weight model can beat a general-purpose frontier model. Our write-up of a $500 RL fine-tune of a 9B open model shows one case, and Bedrock, Fireworks and Vertex all run RL fine-tuning as a service.

FAQ#

When does OpenAI fine-tuning shut down?#

There are three dates. New fine-tuning jobs end for active existing customers on January 6, 2027. o4-mini and GPT-4.1 nano fine-tunes stop working on October 23, 2026. Fine-tunes on GPT-4o, GPT-4o mini, GPT-4.1 and GPT-4.1 mini stop working on January 14, 2027, when their base models shut down.

Will my fine-tuned model keep working after January 6, 2027?#

Yes, until its base model shuts down. OpenAI says "Inference on fine-tuned models will be disabled only when the underlying base model is deprecated." For every base model OpenAI still lists as tunable, that is January 14, 2027 at the latest.

Can I download my OpenAI fine-tuned model weights?#

OpenAI's fine-tuning guides that we read do not state a way to download weights. To move, plan to retrain on another provider from your own training data.

Which providers still offer fine-tuning in October 2026?#

From our directory: Microsoft Foundry (OpenAI models and open models), Vertex AI (Gemini and open models), Amazon Bedrock (Nova, Llama, gpt-oss and Qwen3), Together AI, Fireworks AI, Nebius Token Factory and Alibaba Cloud Model Studio. Cohere's fine-tuning is deprecated, Mistral's is marked deprecated, and Hugging Face AutoTrain is no longer maintained. The fine-tuning directory has every method, price and quote.

Can I fine-tune Claude?#

Not through the Claude API, whose pricing page lists no fine-tuning. Anthropic's cookbook covers Claude 3 Haiku on Amazon Bedrock, but per AWS's lifecycle page that model reached end of life in us-west-2, the Region Bedrock lists for it, on September 10, 2026.

Continue Reading#

• Migrating Off Retired GPT Models in 2026: A Working Checklist - the eval-before-switch process for base models, which is step 4 above in more depth.
• GPT-5.6 Sol, Terra, and Luna: What Each Model Is For - the family OpenAI names as the replacement for the October 23 fine-tunes.
• Bedrock Claude 3.x Shutdown on November 13: Which Claude Model to Move To - the Claude 3 Haiku lifecycle behind the Bedrock caveat above.
• Distilling an LLM on One GPU - the do-it-yourself version of distillation, on an open-weight student.
• Prompt Caching Across AI Providers: Prices, TTLs and Setup - setup code for the cached-prefix approach that can replace a simple fine-tune.

Sources#

Every figure above is a verbatim quote recorded in our fine-tuning dataset or our deprecations tracker, read from these pages:

• OpenAI: deprecations, pricing, supervised fine-tuning, direct preference optimization, reinforcement fine-tuning, vision fine-tuning
• Microsoft: Foundry fine-tuning overview, fine-tuning how-to, Azure OpenAI pricing, model retirement schedule
• Google: Gemini API model tuning, Vertex supervised tuning, Vertex pricing, Vertex model versions
• AWS: custom models, fine-tuning, reinforcement fine-tuning, distillation, distillation prerequisites, Bedrock pricing, model lifecycle
• Anthropic: fine-tuning on Bedrock cookbook, Claude API pricing
• Together AI: pricing, fine-tuning overview, supported models
• Fireworks AI: pricing, fine-tuning models, models
• Nebius: post-training overview, models, how to fine-tune
• Alibaba Cloud: model training overview
• Cohere fine-tuning, Mistral fine-tuning (deprecated), Hugging Face AutoTrain, AutoTrain cost

Fine-tuning data verified October 18, 2026. OpenAI deprecation dates verified October 17, 2026.