Fine-Tuning in 2026: OpenAI and Google Availability

Examine self-serve fine-tuning status across OpenAI and Google in 2026, including sunset deadlines, API schemas, dataset rules, and alternative paths.

By ailog · Published Nov 1, 2026 · 5 min read · How we work
In short
  • OpenAI is winding down self-serve fine-tuning: closed to new users, existing eligible users can run training jobs until January 6, 2027.
  • OpenAI base models and their fine-tuned counterparts shut down on October 23, 2026 (gpt-4.1-nano) and January 14, 2027 (gpt-4.1 and gpt-4.1-mini).
  • Google provides zero fine-tuning models in the standard Gemini API or Google AI Studio; tuning is restricted to Gemini Enterprise Agent Platform.
  • When self-serve fine-tuning is unavailable, teams can evaluate prompt engineering, prompt caching, or base models across different cost tiers.

State of Self-Serve Fine-Tuning Across Major Providers

Self-serve supervised fine-tuning options have shifted significantly across major model APIs. Rather than expanding access, both OpenAI and Google have either restricted or dismantled their standard developer-facing fine-tuning pipelines.

Provider Platform Current Status Supported Models Key Shutdown / Deprecation Dates
OpenAI Fine-Tuning API Winding down; no access for new users; training allowed for eligible users gpt-4.1-2025-04-14, gpt-4.1-mini-2025-04-14, gpt-4.1-nano-2025-04-14 Jobs allowed until January 6, 2027. gpt-4.1-nano shuts down October 23, 2026; gpt-4.1 and mini shut down January 14, 2027.
Google Gemini API / AI Studio Unavailable None (Gemini 1.5 Flash-001 deprecated May 2025) Deprecated May 2025; enterprise users directed to Gemini Enterprise Agent Platform.

For developers planning production infrastructure, these constraints mean that building workflows dependent on self-serve fine-tuning now carries hard lifecycle limits or requires migrating toward alternative platforms.

OpenAI Fine-Tuning: Platform Wind-Down and Lifecycles

OpenAI has officially announced that it is winding down its fine-tuning platform. The service is no longer open to new registrations. For existing eligible accounts, training jobs can still be scheduled until January 6, 2027, provided the underlying base model remains active.

Fine-tuned models remain operational for inference only until their corresponding base model reaches its shutdown date. Developers must track two specific deprecation windows:

  • gpt-4.1-nano-2025-04-14 and all fine-tuned models derived from it shut down on October 23, 2026.
  • gpt-4.1-2025-04-14 and gpt-4.1-mini-2025-04-14 along with their fine-tuned versions shut down on January 14, 2027.

For details on model lifecycles and transitions across platforms, see the guide on model deprecations across Anthropic, OpenAI, and Google.

OpenAI Supervised Fine-Tuning Implementation and Schemas

OpenAI supervised fine-tuning relies on datasets uploaded in JSONL format, adhering to the chat completions structure. Each line corresponds to a distinct training sample with prompt and target assistant content, including function calling specifications if required.

{"messages":[{"role":"user","content":"What is the weather in San Francisco?"},{"role":"assistant","tool_calls":[{"id":"call_id","type":"function","function":{"name":"get_current_weather","arguments":"{\"location\": \"San Francisco, USA\", \"format\": \"celsius\"}"}}]}],"parallel_tool_calls":false,"tools":[{"type":"function","function":{"name":"get_current_weather","description":"Get the current weather","parameters":{"type":"object","properties":{"location":{"type":"string","description":"The city and country, eg. San Francisco, USA"},"format":{"type":"string","enum":["celsius","fahrenheit"]}},"required":["location","format"]}}}]}

The file requires a minimum of 10 examples. Documentation highlights that 50 to 100 well-crafted demonstrations often yield observable improvements for tasks such as classification, nuanced translation, output formatting, or correcting instruction-following failures. Once the JSONL file is prepared, it is uploaded using the Files API with purpose="fine-tune".

curl https://api.openai.com/v1/files \
  -H "Authorization: Bearer $OPENAI_API_KEY" \
  -F purpose="fine-tune" \
  -F file="@mydata.jsonl"

Using the returned file identifier, developers create the training job by calling the fine-tuning endpoint. The default customization method is supervised fine-tuning.

curl https://api.openai.com/v1/fine_tuning/jobs \
  -H "Content-Type: application/json" \
  -H "Authorization: Bearer $OPENAI_API_KEY" \
  -d '{
    "training_file": "file-RCnFCYRhFDcq1aHxiYkBHw",
    "model": "gpt-4.1-nano-2025-04-14"
  }'

Upon completion, the response includes a fine_tuned_model identifier (such as ft:gpt-4.1-nano-2025-04-14:openai::BTz2REMH). This ID is passed directly to the Responses or Chat Completions APIs. OpenAI also generates full model checkpoints at the end of each training epoch to assist when overfitting occurs.

Google Gemini: Developer API Status and Enterprise Route

In the standard Google developer ecosystem, self-serve fine-tuning is no longer available. Google eliminated support inside the Gemini API and AI Studio following the deprecation of Gemini 1.5 Flash-001 in May 2025. There is currently no active model in the Gemini API or AI Studio supporting fine-tuning.

Google directs teams that require model tuning to the Gemini Enterprise Agent Platform. Google has stated it has no immediate plans to restore fine-tuning to the standard Gemini API, directing developer feedback to its community forum.

Teams evaluating Google models for enterprise deployment should also review data handling terms discussed in API data retention and training policies across Anthropic, OpenAI, and Google.

Alternatives to Custom Fine-Tuning in 2026

Because self-serve fine-tuning is deprecated on Google and sunsetting on OpenAI, teams must consider alternative ways to reach desired output quality and cost performance.

  • Prompt optimization and few-shot examples: Supplying explicit instructions and structured formats often achieves results comparable to fine-tuning on small datasets.
  • Model distillation: Teams can refine prompts on larger models (such as gpt-4.1) and capture outputs—retained for 30 days by default in the OpenAI Responses API—to train smaller models while eligible.
  • Structured output enforcement: For strict schema adherence, use native features described in structured outputs across Claude, GPT, and Gemini APIs.
  • Leveraging prompt caching: Caching long context or repeated system prompts significantly lowers inference costs, detailed in prompt caching explained for Claude, GPT, and Gemini.

When switching away from fine-tuned architectures, teams can select low-cost production base models. On the OpenAI side, options include GPT-6 Luna at $0.1 per 1M input tokens ($0.01 cached) and $0.5 per 1M output tokens, GPT-5.4 nano at $0.2 per 1M input tokens ($0.02 cached) and $1.25 per 1M output tokens, or GPT-6.1 Sol at $2 per 1M input tokens ($0.1 cached) and $10 per 1M output tokens.

On Google Gemini, standard alternatives include Gemini 2.5 Flash-Lite at $0.1 per 1M input tokens ($0.01 cached) and $0.4 per 1M output tokens, Gemini 3.1 Flash-Lite at $0.25 per 1M input tokens ($0.025 cached) and $1.5 per 1M output tokens, or Gemini 3.8 Flash at $0.75 per 1M input tokens ($0.075 cached) and $3.75 per 1M output tokens. Claude alternatives include Claude Haiku 4.5 at $1 per 1M input tokens ($0.1 cached) and $5 per 1M output tokens, and Claude Sonnet 5.5 at $2 per 1M input tokens ($0.2 cached) and $10 per 1M output tokens. Calculate potential project expenses using the LLM API cost calculator or explore full rates on the LLM API pricing table.

Implementation and Migration Checklist

Before committing engineering effort to model customization or migrating away from older fine-tuned checkpoints, review the following steps:

  • Verify OpenAI account eligibility: confirm your organization has existing access before designing training jobs.
  • Map base model expiration dates: track the October 23, 2026 cutoff for gpt-4.1-nano-2025-04-14 and January 14, 2027 for gpt-4.1-2025-04-14 and gpt-4.1-mini-2025-04-14.
  • Establish automated evals before training: verify baseline performance using a holdout evaluation dataset before launching fine-tuning jobs.
  • Ensure minimum data volume: maintain at least 10 lines of JSONL data formatted to chat completions specifications, testing with 50 demonstrations initially.
  • Assess Google platform requirements: determine if enterprise needs warrant the Gemini Enterprise Agent Platform given zero support in Gemini API and AI Studio.
  • Benchmark replacement base models: evaluate cost-effective models like GPT-6 Luna, Gemini 2.5 Flash-Lite, or Claude Haiku 4.5 using prompt caching and structured outputs.
Sources
  1. Supervised fine-tuning
  2. Fine-tuning with the Gemini API

Related