Should enterprises still fine-tune models in 2026? OpenAI closure, Google training prices, and RAG decisions

Direct answer: the first 2026 fine-tuning fact is not price but availability: OpenAI states that it is winding down the fine-tuning platform: it is no longer accessible to new users, existing users can create training jobs for the coming months, and fine-tuned models stay available for inference until their base models are deprecated. Google Gemini tuning bills by training tokens/count: Gemini 3.5 Flash is about USD 10 per million training tokens, 3.5 Flash Lite about USD 3, and 2.5 Pro about USD 25; open models range from about USD 0.28–6.83 per million. Training 10 million tokens may cost only USD 30–250, but data preparation, evaluation, inference premium, and retraining dominate TCO. Use RAG for fresh knowledge; consider fine-tuning only for stable format/style/classification with labelled data.

Re-checked 31 October 2026: several models in the original table can no longer be bought. OpenAI's o4-mini, ft-o4-mini and gpt-4.1-nano shut down on 2026-10-23, existing customers lose new fine-tuning jobs on 2027-01-06, and Google's Gemini 2.5 Pro / Flash / Flash-Lite retired on 2026-10-20. The pricing page still lists those training prices, but “price visible” is not “available now”, so the table below adds an official lifecycle column and an availability column; historical prices are kept for comparison only.

This guide is for enterprise AI, product, data, platform, and procurement teams evaluating AI applications. Data cut-off: 31 October 2026; prices come from official OpenAI and Google Cloud pages. AWS Bedrock/Alibaba Cloud tuning prices are not included in the numeric comparison. All prices exclude tax; USD and RMB are not converted and no ranking is implied. This is public-source research and an editorial model, not a quote, and it does not promise accuracy, savings, or a go-live date.

How can enterprises buy fine-tuning in 2026?

OpenAI is closed to new fine-tuning users, and existing active customers lose new training jobs on 2027-01-06 at the latest. Google bills training tokens/count. OpenAI o4-mini/nano and Gemini 2.5 have already shut down or retired. Check the lifecycle before comparing price.

Official prices checked against official lifecycle pages on 31 October 2026. The pricing page still lists training prices for models that have shut down or retired, so “price visible” is not “available now”; availability follows the lifecycle page and the console. Only o4-mini is explicitly listed per training hour; the other OpenAI training figures are shown as the official training price and the invoice defines the unit. Inference bills by token. Google lists Gemini per 1,000 training count and open models per 1,000,000 count, estimated here on a training-token basis. Official footnote: the 1.5× regional price applies only to tuned endpoints of Gemini 3 and later; older Gemini models keep base-model pricing.

What do 1M, 10M, and 100M training tokens cost?

Gemini 3.5 Flash is about USD 10/100/1,000; Flash Lite 3/30/300; 2.5 Pro 25/250/2,500 (historical prices — the model retired on 2026-10-20); small open models are lower.

Editorial model: training tokens × per-million price; excludes tax, data preparation, evaluation, inference, hosting, retraining, and migration. Google lists per 1,000 training count; this table estimates 1 count = 1 token; actual billing follows the official invoice. 100M tokens is a sizing example, not a recommended dataset size. The two Gemini 2.5 rows are historical prices (the models have retired); o4-mini and nano are excluded because their base models have shut down.

Why is training not the total cost?

Fine-tuned inference may cost more, data must be labelled, evaluation/regression is continuous, and versions drift. A USD 100 training job can be dwarfed by inference and data labour.

Editorial guidance, not an accounting taxonomy. Follow corporate finance policy, vendor contracts, and actual invoices.

How much more does fine-tuned inference cost?

OpenAI fine-tuned inference is higher than base: o4-mini about 3.6×, gpt-4.1 about 1.5×, mini/nano about 2×; Google tuned endpoints match base in the global region; from Gemini 3 the non-global region is 1.5×, while older Gemini models stay at base pricing.

Official list prices, USD per million tokens, excluding tax; actual bills depend on region, model version, and contract. OpenAI base prices are from the same official page; the fine-tuned premium scales with monthly volume. o4-mini and gpt-4.1-nano shut down on 2026-10-23; their rows are kept only to show the premium multiple.

Fine-tuning, RAG, or prompting?

Use RAG for new/frequently changing knowledge and citations; consider fine-tuning for fixed format/style/classification with stable labels and meaningful volume.

Editorial guidance, not a vendor ranking. Compare base + RAG/prompting with fine-tuning on the same test set; do not launch without a significant gain.

What should procurement and acceptance measure?

Measure independent-test gain, regression, hallucination, latency, cost per successful task, drift, data rights, and exit — not training loss or a demo.

There is no universal “gain worth fine-tuning” threshold; compare by task risk, volume, and alternatives. High-risk scenarios need human review.

Who should read this — and who should not?

For teams with stable tasks/labelled data and meaningful volume. Not for teams trying to inject fresh knowledge, lacking a test set, or with unclear data rights.

Read if: you are an AI lead, product/algorithm, data, platform, or procurement team; you are comparing RAG, prompting, and fine-tuning; you need to assess training tokens/hours, inference premium, and data/eval cost. Skip if: you want to inject the latest policy/product knowledge via fine-tuning; you have no independent test set; training-data copyright/PII is unclear; volume is tiny and the project is a demo. Supports: availability checks, training cost, inference premium, TCO, fine-tuning/RAG decisions, and acceptance metrics.

Lingyi does not rank model vendors or promise that fine-tuning improves results. If you need to combine RAG/fine-tuning/gateway/permissions/audit into an enterprise AI base, see LingHarness; to validate one task, start with a prompt/RAG baseline. To scope data, evaluation, and budget, talk to us.

See our separate notes on RAG knowledge-base cost, AI acceptance evaluation sets, and private-deployment TCO; for vendor shutdown and migration cost, see AI vendor shutdown risk and exit clauses. This article only covers fine-tuning availability, training/inference cost, and RAG decisions.

FAQ

Can new users still start OpenAI fine-tuning?

Not as a new project. The official deprecation page states: from 2026-05-07, organizations that have never run fine-tuning cannot create training jobs; from 2026-07-02, organizations that have not run fine-tuned inference in the past 60 days cannot either; from 2027-01-06, active existing customers also lose new training jobs. Fine-tuned models keep inference only until their base model shuts down — and o4-mini plus ft-o4-mini went on 2026-10-23. New projects should evaluate Google tuning, open models, or RAG/prompting.

What does 10 million training tokens cost?

At Google public prices: Gemini 3.5 Flash about USD 100 and Flash Lite 30; 2.5 Pro about 250 and 2.5 Flash about 50 (both retired on 2026-10-20, historical only); Gemma 3 4B about 11.40 and Llama 3.1 8B about 6.70. Add data, evaluation, inference, and hosting; the table is an editorial model, not a quote.

Why is training cheap but total cost high?

Training is one-off; data labelling, independent test sets, regression, inference premium, endpoints/regions, retraining, and migration recur. OpenAI fine-tuned inference can be 1.5–3.6× the base; the premium grows with volume.

Can fine-tuning replace RAG?

No. Fine-tuning suits fixed format/style/classification/behaviour, not frequently changing knowledge; RAG suits updatable, citable knowledge. Most enterprise cases should establish a RAG/prompting baseline before deciding whether fine-tuning gives a significant independent-test gain.

Does Google bill tuning by token or hour?

Google lists Gemini per 1,000 training count and open models per 1,000,000 count (estimated here on a training-token basis); tuned endpoints match base in the global region, and from Gemini 3 the non-global region is 1.5×, while older Gemini models are unchanged. Note the retirement dates: Gemini 2.5 Pro/Flash/Flash-Lite retired on 2026-10-20; 3.5 Flash is 2027-05-19 or later and 3.5 Flash Lite 2027-07-21 or later. Actual billing unit, region, and model version follow the console/invoice; OpenAI lists o4-mini training per hour and inference per token.

Are the models in the table still available?

Judge by the official lifecycle page, not the pricing page: o4-mini, ft-o4-mini and gpt-4.1-nano shut down on 2026-10-23, the three Gemini 2.5 models retired on 2026-10-20, and gpt-4.1 / 4.1-mini shut down on 2027-01-14. The pricing page keeps their prices for historical comparison only; new projects should choose among 3.5 Flash, 3.5 Flash Lite, 3.1 Flash Lite or open models, and confirm availability in the console.

How should we accept fine-tuning results?

Compare fine-tuning with base+RAG/prompting on an independent test set, by task: accuracy, regression, hallucination, latency, cost per successful task, drift, and data rights. Do not rely on training loss, a vendor demo, or “larger model”; do not launch without a significant gain.

Method and limitations

This article uses only official pricing/product pages accessible on 31 October 2026. It does not use private contracts, customer invoices, or vendor internal data. Prices, models, training billing units, availability, and terms can change. The article has no actual enterprise agreement, tax, FX, data-labelling, evaluation, or inference data, so every cost is an editorial model, not a quote. Google lists Gemini per 1,000 training count and open models per 1,000,000 count; this article estimates a training-token basis. OpenAI fine-tuning is winding down; the existing-user window and the inference-retention period follow the official page and deprecation timeline. This article re-checked OpenAI Deprecations and Google Model versions and lifecycle on 31 October 2026: because the pricing page still lists training prices for models that have shut down or retired, availability is judged by the lifecycle page and prices are kept for historical comparison only. AWS/Alibaba Cloud tuning prices are not included.

Sources and basis

Data cut-off: 31 October 2026. The following are official pricing/document pages; prices, models, availability, and terms follow the original page and contract.

1. OpenAI, API Pricing: platform.openai.com/docs/pricing. Fine-tuning platform wind-down (closed to new users, existing users for months, fine-tuned models stay for inference), o4-mini data-sharing discount, gpt-4.1 training prices, and fine-tuned inference token prices.
2. Google Cloud, Vertex AI generative AI pricing: cloud.google.com/vertex-ai/generative-ai/pricing. Gemini/open-model supervised tuning per 1,000 training count and tuned-endpoint regional pricing.
3. Google Cloud, Supervised fine-tuning documentation: cloud.google.com/vertex-ai/generative-ai/docs/models/tuning/supervised-tuning. Data preparation, training, and evaluation process entry.
4. OpenAI, Deprecations: developers.openai.com/api/docs/deprecations. o4-mini / ft-o4-mini / gpt-4.1-nano shut down 2026-10-23; gpt-4.1 series 2027-01-14; from 2027-01-06 existing customers lose new fine-tuning jobs.
5. Google Cloud, Model versions and lifecycle: cloud.google.com/vertex-ai/generative-ai/docs/learn/model-versions. Gemini 2.5 Pro/Flash/Flash-Lite retired 2026-10-20; 3.5 Flash 2027-05-19 or later; 3.5 Flash Lite 2027-07-21 or later.