30 October 2026 · Site Methodology & Pricing Notes · 6 min read
Nuestro Calculadora de costos de ajuste has a dropdown with five models in it today. This morning it had six, and one of them -- "Google · Gemini 2.5 Flash," priced at $3.00 per million tokens to train -- described a product that Google stopped selling in May 2025. Not a price change. Not a renamed tier. The feature itself doesn't exist. I only found out because I went looking for a new blog topic and decided to actually re-verify every provider on that page instead of assuming the dropdown was still true.
Straight from Google's current fine-tuning docs, checked today: "With the deprecation of Gemini 1.5 Flash-001 in May 2025, we no longer have a model available which supports fine-tuning in the Gemini API or AI Studio." Not "fewer models." Zero. That line has apparently been sitting there for about 17 months while our calculator kept offering a fine-tuned Gemini 2.5 Flash as if you could buy it. You can't. Tuning still technically exists as a product, just not this one -- it's sold separately through the Gemini Enterprise Agent Platform, which is sales-gated, not something you self-serve from the API console most developers actually use.
OpenAI's current guide is more direct about its own shutdown than I expected: "OpenAI is winding down the fine-tuning platform. The platform is no longer accessible to new users." Existing users can keep submitting jobs "for the coming months," then that's it. And look at what's actually still fine-tunable right now -- every model on the list is a dated 2024-2025 snapshot:
| Modelo | Método | Training price |
|---|---|---|
| gpt-4.1-2025-04-14 | SFT / DPO | $25.00 / 1M tokens |
| gpt-4.1-mini-2025-04-14 | SFT / DPO | $5.00 / 1M tokens |
| gpt-4.1-nano-2025-04-14 | SFT / DPO | $1.50 / 1M tokens |
| gpt-4o-2024-08-06 | Vision fine-tuning | $25.00 / 1M tokens |
| o4-mini-2025-04-16 | Reinforcement (RFT) | $100.00 / hour |
No GPT-5. No current flagship anywhere on that menu. If you've been assuming you could fine-tune whatever OpenAI's best model happens to be this month, the honest answer has been "no" for a while -- you've only ever been able to fine-tune last year's mid-tier snapshot, and soon not even that if you're a new account.
This was the one I expected to be wrong about, and wasn't. Anthropic has never exposed self-serve fine-tuning through its own public API. The only supported route is supervised fine-tuning of Claude 3 Haiku -- two model generations behind Anthropic's current line -- through Amazon Bedrock in the us-west-2 region specifically. That's not a missing feature that got removed; it's a feature that was never built for the API most people reach for first.
Put the three side by side and the pattern is blunt:
| Proveedor | Fine-tune via own API today? | Newest fine-tunable model |
|---|---|---|
| AbiertoAI | Yes, but closed to new users, winding down | GPT-4.1 family (Apr 2025 snapshot) |
| Google Géminis | No model available since May 2025 | Gemini 1.5 Flash-001 (deprecated) |
| antrópico | No, never offered on the public API | Claude 3 Haiku, Bedrock-only |
Ask "can I fine-tune today's flagship model through the vendor's own API" and all three frontier labs answer no, for three completely different reasons and on three different timelines. That's not a coincidence worth a shrug -- it's three separate business decisions landing in the same place.
I pulled the Gemini row out of both the Calculadora de costos de ajuste y el Fine-Tuning vs Prompting Calculator. It wasn't a wrong number -- a wrong number would have been easy to catch by just double-checking the price. This was a row in a dropdown that quietly kept describing a live purchase option for something that stopped being purchasable a year and a half ago. Our other audits this year have mostly caught formulas computing the wrong thing. This one wasn't a formula at all. It was a feature that disappeared out from under a page nobody had re-checked since it shipped.
New to fine-tuning cost math in general? Start with the guías gratuitas de costos de API.
Despliéguelo usted mismo: DigitalOcean — $200 de crédito gratis ↗ · VPS Hostinger ↗
Checked directly against ai.google.dev/gemini-api/docs/model-tuning, developers.openai.com/api/docs/guides/model-optimization and developers.openai.com/api/docs/pricing, and public statements on Anthropic's supported Bedrock fine-tuning path, 2026-10-30. Reference figures, not live quotes -- confirm current availability on each provider's own docs before planning a fine-tuning project around any of these.