OpenAI Deprecates Finetuning APIs, Forcing Shift to Instruction‑Tuning
Migrate your models away from OpenAI finetuning to alternative fine‑tuning or instruction‑tuning approaches.
Migrate your models away from OpenAI finetuning to alternative fine‑tuning or instruction‑tuning approaches.
Summary
OpenAI has deprecated its finetuning APIs, a feature that had enabled developers to achieve O1‑level performance at GPT‑4‑price points.
The deprecation marks a significant shift for the majority of the AI engineering community, which had relied on finetuning for model customization. Anthropic is poised to raise its valuation, while other tools such as Cursor, Cognition, and Sora are stepping up their open‑model RLFT and usage. The move is part of the 2026 Side Quest wave, which also saw the deprecation of Sora and other fine‑tuning services. Despite the GPU crunch narrative, the industry’s reliance on finetuning was already high, with many teams using it for high‑quality, low‑latency real‑time support. Alternatives include instruction‑tuning, RLHF, and custom ASIC solutions that can handle longer prompts without finetuning. The deprecation forces developers to rethink their model pipelines and explore new fine‑tuning or instruction‑tuning strategies.
Key changes
- OpenAI deprecates finetuning APIs
- Finetuning enabled O1 performance at GPT‑4 price
- Anthropic poised to raise valuation
- Cursor, Cognition, Sora stepping up open‑model RLFT
- 2026 Side Quest wave includes deprecation of Sora
- Industry reliance on finetuning was high
- Alternatives: instruction‑tuning, RLHF, custom ASIC solutions