OpenAI Fine-Tuning API

source

closed

training-code

closed

runs-on

OpenAI-platform-only

license

Proprietary(paid API service)

Proprietary managed fine-tuning service, no source, runs only on OpenAI's platform against OpenAI's closed models.

• https://openai.com/policies/services-agreement/ recorded 2026-08-08

  Section 9.1 Reservation of Rights: "Customer obtains only a limited right to use the Services, and no ownership rights are transferred to Customer or its End Users under this Agreement." The use restrictions bar Customer from "Reverse Engineer any aspect of the Services or the systems used to provide the Services", and the definition of Reverse Engineer covers any attempt to "otherwise attempt to discover the source code or underlying components of the Services". The Fees section requires Customer to pay "the applicable Fees in the currency and pursuant to the payment terms on the Order Form", and the Permitted Exception definition reaches "fine tune or customize models provided as part of OpenAI"'s fine-tuning or other Services set forth on the Pricing Page, which ties this Agreement to this product specifically. Header: Updated December 1, 2025; Effective January 1, 2026.

• https://developers.openai.com/api/docs/guides/model-optimization recorded 2026-08-08

  The fine-tuning workflow is entirely OpenAI-hosted: "Collect a dataset of examples to use as training data", "Upload that dataset to OpenAI, formatted in JSONL", then create a job — and "you can create fine-tuned models either in the" dashboard "or" "with the API". No repository, package, or self-hostable runtime is offered anywhere on the page. It also points at the pricing page for "how fine-tuned model training and usage are billed", and carries the banner "OpenAI is winding down the fine-tuning platform" with the platform "no longer\naccessible to new users".

• https://developers.openai.com/api/docs/deprecations recorded 2026-08-08

  Under "Update to OpenAI"'s self-serve fine-tuning, a three-row table dated from the May 7th 2026 notification. May 7, 2026: "Creating fine-tuning jobs or training is not available to organizations that have not previously run fine-tuning." July 2, 2026: "Creating 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, 2027: "Active existing customers will no longer be able to create new fine-tuning jobs on this date. Inference on fine-tuned models will be disabled only when the underlying base model is deprecated." Cited as lifecycle corroboration; it settles no scored openness dimension on its own.

No standalone usage figure is disclosed for the fine-tuning API -- no developer count, no jobs run, no tokens trained. The OpenAI platform broadly serves ~1B+ weekly ChatGPT users, but that surface is not the fine-tuning API and cannot honestly be attributed to it, so no level is assigned rather than one borrowed from the umbrella product.

• https://developers.openai.com/api/docs/pricing recorded 2026-08-08

  A "Finetuning" section, "Prices per 1M tokens.", carrying the wind-down banner and a Standard/Batch rate table. The visible rows are o4-mini-2025-04-16 at "$100.00 / hour" training; the collapsed rows in the body carry training rates for gpt-4.1-2025-04-14 (25), gpt-4.1-mini-2025-04-14 (5), gpt-4.1-nano-2025-04-14 (1.5), gpt-4o-2024-08-06 (25), gpt-4o-mini-2024-07-18 (3), and legacy gpt-3.5-turbo (8) and davinci-002 (6), under the column order Model / Training / Input / Cached input / Output. Rates only: the page publishes no count of users, organizations, jobs, or tokens.

• https://developers.openai.com/api/docs/guides/model-optimization recorded 2026-08-08

  OpenAI's own product documentation for the fine-tuning platform: the optimization workflow, the four supported methods and their base snapshots, and a pointer to the pricing page for billing. It describes who may use the service ("no longer\naccessible to new users") but publishes no figure for how many do.

• https://openai.com/api/pricing/ recorded 2026-08-08

  Redirects (final_url https://openai.com/business/pricing/) to "Business Pricing | OpenAI", a ChatGPT Business ($20 / user / month) and Enterprise ("Custom pricing") plan-comparison page. It carries no API token rates and no fine-tuning content at all — a search of the extracted text for "fine-tun" returns zero hits. Recorded to document where the canonical API pricing URL now lands; the rate card lives at developers.openai.com/api/docs/pricing.

Strong managed method coverage over frontier base models: supervised fine-tuning, vision fine-tuning, DPO preference tuning and reinforcement fine-tuning for reasoning models, applied to the gpt-4.1 family, gpt-4o and o4-mini, with data validation, training, hosting and evaluation all handled end to end. It is a black box, though: no parallelism, precision or scale controls are exposed and no training benchmark stands behind it. The score reflects that method breadth and base-model quality, and stops short of the top band because the user cannot control training scale or internals.

• https://developers.openai.com/api/docs/guides/model-optimization recorded 2026-08-13

  "These are the fine-tuning methods supported in the OpenAI platform today." — a four-row table: "Supervised fine-tuning (SFT)" with gpt-4.1-2025-04-14, gpt-4.1-mini-2025-04-14, gpt-4.1-nano-2025-04-14; "Vision fine-tuning" with gpt-4o-2024-08-06; "Direct preference optimization" (DPO) with the same gpt-4.1 trio; and "Reinforcement fine-tuning (RFT)", marked "Reasoning models only", with o4-mini-2025-04-16. The surrounding text describes the managed loop end to end (upload JSONL, create the job, define a grader for RFT, evaluate) and exposes no parallelism, precision, or training-scale controls.

• https://developers.openai.com/api/docs/deprecations recorded 2026-08-08

  The "2026-04-22: Legacy GPT model snapshots" table lists "October 23, 2026" shutdowns for "o4-mini-2025-04-16" (replacement gpt-5.6-terra) and gpt-4.1-nano-2025-04-14 (replacement gpt-5.6-luna), and a following table "We are also removing fine-tuned versions as below" lists ft-o4-mini-2025-04-16 and ft-gpt-4.1-nano-2025-04-14 on the same date. o4-mini-2025-04-16 is the sole base the guide offers for reinforcement fine-tuning.