See what your AI agent really costs

How Much Does Fine-Tuning Cost, and Is It Still Worth It?

Updated 2026-10-25. Checked against official provider docs and pricing pages on 2026-10-25.

For years, the standard advice on cutting LLM costs was "fine-tune a smaller model." In 2026 that advice needs a big caveat. OpenAI is winding down its self-serve fine-tuning platform. Before you budget for training runs, check whether you can still create one, and how long the resulting model will stay online.

This guide covers what fine-tuning officially costs, the shutdown timeline, what other providers offer, and the cheaper alternatives to try first.

OpenAI's fine-tuning wind-down: the official timeline

OpenAI's deprecations page has a section titled "Update to OpenAI's self-serve fine-tuning." It says OpenAI notified developers on May 7, 2026, and lays out this schedule:

DateWhat changes (per OpenAI)
May 7, 2026Creating fine-tuning jobs or training is not available to organizations that have not previously run fine-tuning.
July 2, 2026Creating fine-tuning jobs is no longer available to organizations that have not run inference on a fine-tuned model in the past 60 days.
Jan 6, 2027Active existing customers will no longer be able to create new fine-tuning jobs. Inference on fine-tuned models will be disabled only when the underlying base model is deprecated.

The fine-tuning guides now open with a banner. Both the supervised and reinforcement fine-tuning guides say: "OpenAI is winding down the fine-tuning platform. The platform is no longer accessible to new users. Eligible existing customers can create training jobs until January 6, 2027, unless the base model shuts down earlier." The banner on the pricing page is vaguer: existing users "will be able to create training jobs for the coming months." The deprecations page and guide banners give the hard date: January 6, 2027.

Your fine-tuned model lives only as long as its base model

That last clause matters most for cost planning. The same deprecations page lists these shutdowns:

So a model you fine-tune today on GPT-4.1 mini has, at most, until January 14, 2027 to earn back its training cost.

What OpenAI fine-tuning costs (official list prices)

These are from OpenAI's pricing page, Finetuning section, retrieved 2026-10-25. Prices are per 1M tokens. Training is per 1M training tokens. Only base models whose fine-tunes are still running are listed.

Base modelTrainingInputCached inputOutput
gpt-4.1-2025-04-14$25.00$3.00$0.75$12.00
gpt-4.1-mini-2025-04-14$5.00$0.80$0.20$3.20
gpt-4o-2024-08-06$25.00$3.75$1.875$15.00
gpt-4o-mini-2024-07-18$3.00$0.30$0.15$1.20

The page also lists batch prices for fine-tuned models, which are lower than standard. Some rows still shown there belong to fine-tunes that shut down on Oct 23, 2026 (including o4-mini's per-hour reinforcement fine-tuning), so they are omitted. The page says that tokens used for model grading in reinforcement fine-tuning "are billed at that model's per-token rate," and that inference discounts are available if you enable data sharing.

Training cost formula

Training cost ≈ training tokens per epoch × epochs × training price ÷ 1,000,000

For example, with an illustrative dataset of 2M tokens trained for 3 epochs on gpt-4.1-mini: 2 × 3 × $5 = $30. Your dataset size and epoch count are your own; OpenAI's fine-tuning guide explains how they're set.

The catch: fine-tuned inference vs newer base models

Training is often the small part. You pay the fine-tuned inference rate on every call after that. Compare with the replacement OpenAI recommends, using prices from TokenTab's models.json (retrieved 2026-10-25):

ModelInput / output per 1M
Fine-tuned gpt-4.1-mini$0.80 / $3.20
Fine-tuned gpt-4o-mini$0.30 / $1.20
GPT-6 Luna (base, no tuning)$0.10 / $0.50

At list prices, an untuned GPT-6 Luna costs less per token than either fine-tuned mini model, and it isn't scheduled to shut down in January. Whether it matches your fine-tuned model's quality is something only your own evals can answer. To see what that per-token gap means per month for your workload, enter your token counts in the TokenTab AI agent cost calculator.

Other providers (official pages only)

Is it still worth it?

Fine-tuning can still make sense if all of these are true:

  1. You're an eligible existing OpenAI fine-tuning customer, or you use an open-weights model you host yourself.
  2. Your base model's shutdown date leaves enough time to recover the training cost.
  3. Your evals show the fine-tuned model clearly beats a newer, cheaper base model with good prompting.

For most new projects, try these alternatives first.

1. Better prompting with examples

Few-shot examples and a tight system prompt often get you the consistent style or format people fine-tune for. The trade-off is more input tokens per call.

2. Prompt caching

If those examples sit in a stable prompt prefix, caching makes them cheap to repeat. For example, GPT-6 Luna cached input is $0.01 per 1M vs $0.10 uncached, and Claude Haiku 5.5 is $0.01 vs $0.10 (models.json). Try TokenTab's prompt caching calculator, and see does prompt caching save money for AI agents?

3. Retrieval (RAG)

If the goal is teaching the model your facts, retrieval keeps knowledge current without retraining. Fine-tuning is generally for behavior and format, not for updating knowledge.

4. Controlling output length

Output is the expensive side. On OpenAI, cap it with max_output_tokens (Responses API) or max_completion_tokens (Chat Completions). On Anthropic, thinking tokens are billed as output tokens, and with manual extended thinking the budget_tokens value is "a target rather than a strict cap"; max_tokens is the hard ceiling. Anthropic has deprecated manual extended thinking on Claude 4.6 models, and Claude 4.7 and later reject it in favor of adaptive thinking (docs).

Sources

Retrieved 2026-10-25.

Get the TokenTab newsletter

Model price changes and agent cost tips, straight to your inbox. Unsubscribe anytime; we only use your email for this newsletter.