Key findings
- The Assistants API shut down on August 26, 2026, a year after its deprecation notice. OpenAI names the Responses and Conversations APIs as the replacement.
- Existing evals become read-only on October 31, 2026, a month before the Evals platform closes. The v1/prompts API with reusable prompt objects closes the same day as Agent Builder.
- Self-serve fine-tuning closed to new organizations on May 7, 2026. Active customers lose new training jobs on January 6, 2027, and fine-tuned models serve until their base models retire.
- Older GPT-5 and o3 snapshots retire on December 11, 2026, and GPT-5.3-Codex, GPT-5.1, and GPT-5.4-nano on April 1, 2027.
- Of the 112 public accounts GILD had read by October 26, 2026, 17 discuss model and API changes and 12 raise them as a concern.
Budget migration work for every OpenAI shutdown on your roadmap, even where OpenAI names a replacement. The next deadline is November 30, 2026, when the Evals platform, Agent Builder, and reusable prompts close.
The schedule
OpenAI gives generally available models at least 6 months' notice before retirement, specialized variants such as Codex and chat variants at least 3 months, and early-release models as little as 2 weeks. GILD checked the dates below against OpenAI's deprecations page on October 25, 2026.
- Legacy GPT-3.5, GPT-4, GPT-4 Turbo, GPT-4.1 nano, gpt-4o-2024-05-13, o1, o1-pro, o3-mini, o4-mini, and gpt-image-1: shut down October 23, 2026; calls fail. Substitutes are GPT-5.6 Sol, Terra, or Luna, and gpt-image-2.5 for images.
- Assistants API: shut down August 26, 2026; replaced by the Responses and Conversations APIs.
- Evals platform: read-only October 31, 2026, shut down November 30, 2026; OpenAI points to Promptfoo as a migration path.
- Agent Builder: shuts down November 30, 2026; replaced by the Agents SDK or ChatGPT Workspace Agents.
- Reusable prompts and v1/prompts: removed on November 30, 2026, with prompt content moving into application code.
- Older GPT-5 and o3 snapshots: retire December 11, 2026; replaced by GPT-5.6 Sol, Terra, and Luna.
- GPT-5.3-Codex, GPT-5.1, and GPT-5.4-nano: retire April 1, 2027; replaced by GPT-6 Sol and GPT-6 Luna.
- GPT-4o (2024-08-06 and 2024-11-20), GPT-4o mini, GPT-4.1, and GPT-4.1 mini: retire January 14, 2027; replaced by GPT-6 Luna.
Fine-tuning is winding down
OpenAI closed fine-tuning to organizations that had never used it on May 7, 2026, and to those without recent fine-tuned inference on July 2, 2026. Active customers can create training jobs until January 6, 2027, and fine-tuned models keep serving until their base models are deprecated.
One builder replaced fine-tuned GPT-4.1-mini with GPT-6 Luna at matching accuracy and competitive cost. GPT-5-mini had been slower and more expensive. Another team found Google and OpenAI retiring the fast, cheap models its fine-tuned workloads needed in favor of slower, costlier reasoning models.
Pricing and caching rules changed too
Priority processing became Fast mode on July 30, 2026. On GPT-5.6 and later, cache writes cost 1.25x the input rate. Reads cost 0.1x, or 0.05x on GPT-6.1 Sol and GPT-6.1 Terra. prompt_cache_options.ttl takes 30 minutes or, since October 20, 2026, 12 hours at a 2.5x write rate on Responses, and explicit breakpoints are new. GPT-5.6 also dropped partial-prefix matching. Ploy's new conversations cached none of a shared 29K-token prefix and paid the write premium on every call. Setting prompt_cache_key per customer workspace fixed it. A layout that cached well on older models can now cost more than no cache.
What the changes cost teams
Teams paid very different migration costs. One builder moved an Assistants MVP to Responses with little work, and another moved every project to the Agents SDK. Definite instead left Assistants for a model-agnostic framework. A developer stopped using stateful model APIs after the Assistants shutdown. A production voice team reverted from gpt-realtime-2.1-mini when it narrated instructions aloud. Another builder noted that GPT-5 mini's replacement, GPT-5.4 mini, costs three times as much per input token.
Methodology
GILD read OpenAI's deprecations, pricing, prompt caching, and Fast mode documentation, checked October 12, 2026, and rechecked the deprecations and prompt caching pages on October 25, 2026. Practitioner counts come from the 112 public accounts GILD collected for the OpenAI API, published between January 31, 2025 and October 15, 2026.
Products in this piece
Sources
- Official documentation
- 9
- Public source
- 9
- OpenAI API docs: Deprecations
- OpenAI API docs: Assistants migration guide
- OpenAI API docs: Migrate to the Responses API
- OpenAI API docs: Evaluation best practices
- OpenAI API docs: Pricing
- OpenAI API docs: Prompt caching
- OpenAI API docs: Changelog
- OpenAI API docs: Fast mode
- OpenAI API docs: Using GPT-6
- Hacker News comment on replacing a fine-tuned GPT-4.1-mini with GPT-6 Luna (Sep 23, 2026)
- Hacker News comment on fast models being deprecated (Jul 10, 2026)
- Ploy: Migrating a production AI agent to GPT-5.6 (Jul 9, 2026)
- Hacker News comment on migrating an MVP from Assistants to Responses (May 21, 2025)
- Hacker News comment on moving projects to the Agents SDK (May 29, 2026)
- Hacker News comment from Definite on leaving the Assistants API (May 21, 2025)
- Hacker News comment on the Assistants API retirement (Sep 11, 2026)
- OpenAI Developer Forum: A production voice agent's regression on gpt-realtime-2.1-mini (Aug 4, 2026)
- Hacker News comment on GPT-5 mini's retirement and pricier replacements (Jun 26, 2026)