OpenAI Multimodal Fine-Tuning Is Shutting Down: What the Notice Says

OpenAI closed multimodal fine-tuning to new users on May 8, 2026.

Existing users got "the coming months" to keep creating training jobs, which is a duration, not a date. Every fine-tuned model already in production stays alive exactly as long as its base model does. Not longer.

Start there, because that last clause decides everything downstream. Per the notice: "All fine-tuned models will remain available for inference until their base models are deprecated."

The memo itself sits inside OpenAI's fine-tuning documentation, on a page saying the enterprise "is winding down the fine-tuning platform," that it is "no longer accessible to new users," and that existing users "will be able to create training jobs for the coming months." Three commitments.

None of them includes a calendar.

I run a one-person automation shop, and my honest read: this was predictable, not a betrayal. Below is what the fine print says, why the frontier roadmap made the shutdown look scheduled. And what to do while the window stays open.

What OpenAI's Multimodal Fine-Tuning Shutdown Actually Says

The page hosting this memo was originally published April 4, 2024, under the title "Introducing improvements to the fine-tuning API and expanding our custom models program." Two years between that announcement and the wind-down language. Same URL, two very different documents.

Three commitments, in order.

No new users. Door shut. No waitlist language, no enterprise exception, nothing. If you lacked fine-tuning platform access before the update, OpenAI multimodal fine-tuning is simply not available to you.

Existing users get "the coming months." A duration dressed up as a commitment.

Nobody schedules a migration against a phrase.

Inference survives until base-model deprecation. The fine-tune has no end date of its own.

The base model's deprecation date is your model's deprecation date. Copy that line into your notes.

One capability still stands in the docs: hyperparameter configuration from the Dashboard, "rather than only through the API or SDK." Dashboard polish on a platform in wind-down reads as maintenance mode, not expansion. When the last shipped improvement is a UI for something newcomers cannot sign up for, the roadmap has already left the building.

How Tool-Use Fine-Tuning Went Legacy

Sequence matters more here than any single memo.

July 23, 2025: a practitioner starts a thread in the OpenAI community forum titled "Finetuning with tool calls and tool responses," thread 1322493. Real users still working through the mechanics of training on tool-call traces. The exact workflow the "newly released capabilities" story was built around.

September 4, 2026: OpenAI's model optimization guide gets an update and now opens by stating it "covers evals and fine-tuning workflows that are being moved into legacy documentation." Official material for evaluating and fine-tuning, retired to an archive.

A DeepWiki index of the OpenAI Cookbook shows the fine-tuning page itself still being touched as recently as September 18, 2026.

That index is the only source for it. Single-sourced, so treat it as directional.

Stack the three moments. Mid-2025, practitioners mid-conversation about tool-call training. May 2026, platform closed to newcomers. September 2026, official workflows headed for legacy status.

Not a feature maturing.

Absorption, more like.

The Frontier Roadmap Called This First

This is where the sourcing thins out.

One release-tracking source, and only that one, reports the new flagship shipped September 3, 2026 with API id gpt-6-astra, succeeding the GPT-5.6 family released in July — Sol as the flagship, Terra as the balanced everyday model, Luna as the cost-efficient tier. No second source in hand. Single-sourced. Hold it loosely.

The structural line in that same report: "OpenAI's flagship now coordinates parallel agents as a reasoning setting." That quote is the part with consequences. If coordinating parallel, tool-using agents is now a base-model setting, then the two things a team would fine-tune for — consistent tool calls and reliable multi-step behavior — turn into prompt configuration. And notice the timing: the flagship shipped September 3, 2026. The guide update declaring fine-tuning workflows legacy landed September 4, 2026. Docs went to legacy the morning after the capability moved into the base model.

Same source calls gpt-6-astra OpenAI's most capable and most aligned model to date, then adds, "The claims are broad." Undercutting itself in the next sentence is honest of it. Superlatives are marketing.

The parallel-agents detail is the load-bearing fact.

Economics sit underneath all of it.

Every prompt bills on OpenAI's meter.

A fine-tune you own is a variant they host while you keep the differentiation. Sunsetting the second thing to concentrate demand on the first is not a mystery.

It is a strategy.

For small teams this cuts both ways. Base models absorbing tool-use skill means less infrastructure to babysit. And less of the product that is actually yours. Like a bakery buying in pre-made dough. Fewer steps to manage, and less of the loaf is yours.

OpenAI Fine-Tuning Deprecation: Your Exit Plan

Four moves, starting now rather than when a deprecation email lands.

Inventory every fine-tune next to its base model. Base deprecation is your deprecation. Write the pairings down so nothing arrives as a surprise.

Spend the remaining training budget this quarter. "The coming months" carries no renewal clause. So list the jobs that genuinely matter and queue them while job creation still works.

Own the evals. The workflows for evals and fine-tuning are moving into legacy documentation. And a benchmark that lives only in their material leaves with it. The eval set is the asset that survives any platform decision.

Nothing else on their side does.

Benchmark prompt-only against the current flagship before training anything. My operating rule on client automations is prompts before weights, always, and a fine-tune only when prompting fails measurably. Weeks like this one are why that rule exists.

If you never had platform access, none of this requires panic.

The people taking real damage are the ones who built product on retraining access they assumed was permanent infrastructure.

Fine-Tuning Shutdown FAQ

Is OpenAI fine-tuning still available?

Partly. The platform is "no longer accessible to new users." Existing users "will be able to create training jobs for the coming months," per the notice.

Inference on models already trained continues.

Will my fine-tuned models stop working?

Not on their own schedule. "All fine-tuned models will remain available for inference until their base models are deprecated." The end date belongs to the base model, never to the fine-tune.

What happens to tool-use fine-tuning?

The official workflows are being retired. OpenAI's model optimization guide, updated September 4, 2026, states it "covers evals and fine-tuning workflows that are being moved into legacy documentation." Community practitioners were still trading notes on tool-call traces as of July 23, 2025. The capability goes to legacy docs while closed to newcomers.

What should I do before the window closes?

Inventory every fine-tune with its base model. Queue the training jobs that matter while job creation stays open. Move your evals somewhere you control. Benchmark prompt-only against the current flagship before committing weights to anything.

What the Fine-Tuning Shutdown Means for Builders

OpenAI multimodal fine-tuning and tool-use training went from announced capability to wind-down notice without ever being stable ground for a business. April 4, 2024, a product announcement.

May 8, 2026, a shutdown memo on the same page.

In between, people built.

The layer a builder can actually own is the base model plus the evals plus the data. Weights you cannot retrain are not an asset. They are a lease, and the landlord holds the demolition date.

Audit what runs on a fine-tune this week, while the timing is still yours.

That will not stay true.

Sources

- OpenAI, fine-tuning documentation — "Introducing improvements to the fine-tuning API and expanding our custom models program" (originally published April 4, 2024; wind-down language per the May 8, 2026 update): https://openai.com/index/introducing-improvements-to-the-fine-tuning-api-and-expanding-our-custom-models-program/ - OpenAI Developers, model optimization guide, updated September 4, 2026: https://developers.openai.com/api/docs/guides/model-optimization - OpenAI Community Forum, "Finetuning with tool calls and tool responses," July 23, 2025: https://community.openai.com/t/finetuning-with-tool-calls-and-tool-responses/1322493 - DeepWiki, index of the OpenAI Cookbook, fine-tuning page (single-sourced; directional): https://deepwiki.com/openai/openai-cookbook/8-fine-tuning - Release-tracking newsletter "Frontier Dispatch, Issue 02," Continuum Resources (single source for the gpt-6-astra and GPT-5.6 family details): https://www.continuumresourcesllc.com/ai-newsletters/frontier-dispatch-issue-02.html