Last reviewed: September 7, 2026.
Quick answer: Which OpenAI APIs and models are shutting down in 2026?
OpenAI has multiple shutdowns remaining in 2026. The most urgent is the Videos API and Sora 2 models on September 24, followed by legacy completions and GPT-3.5 models on September 28. A larger group of GPT-4, GPT-4o, o1, o3-mini, and fine-tuned snapshots shuts down October 23. Reusable prompts, the Evals platform, and Agent Builder end November 30. Older GPT Image models end December 1, and older GPT-5 and o3 snapshots end December 11.
Do not treat every migration as a model-ID replacement. Sora 2 currently has no listed replacement; reusable prompts must move into application code; Evals needs an external evaluation system; Agent Builder moves to the Agents SDK or ChatGPT Workspace Agents. Inventory model IDs and product endpoints now, add contract tests, and deploy replacements behind configuration flags before each deadline.
What is the OpenAI API deprecation calendar for the rest of 2026?
| Shutdown date | Models or product | Recommended action |
|---|---|---|
| Sep 24 | Videos API, sora-2, sora-2-pro, listed snapshots |
No replacement listed; remove dependency or redesign video flow |
| Sep 28 | gpt-3.5-turbo-instruct, babbage-002, davinci-002, gpt-3.5-turbo-1106 |
Move to gpt-5.6-terra |
| Oct 23 | Selected GPT-3.5, GPT-4, GPT-4o, o1, o3-mini, GPT Image, and fine-tuned snapshots | Use the official per-model replacements and re-evaluate outputs |
| Oct 31 | OpenAI Evals | Existing evals become read-only |
| Nov 30 | /v1/prompts, reusable prompts, Evals dashboard/API, Agent Builder |
Move prompts to code; move evals; use Agents SDK or Workspace Agents |
| Dec 1 | gpt-image-1-mini, gpt-image-1.5, chatgpt-image-latest |
Move to gpt-image-2 |
| Dec 11 | Older GPT-5 and o3 snapshots | Move to GPT-5.6 Sol, Terra, or Luna as specified |
This table is a planning summary, not a substitute for OpenAI’s live deprecations page. OpenAI can add notices, so check the official source during every release cycle.
What should Sora 2 and Videos API developers do before September 24?
OpenAI lists the Videos API, sora-2, sora-2-pro, and multiple dated snapshots for shutdown on September 24, 2026. The official table does not provide a replacement.
That changes the migration plan. Do not silently point production to an assumed successor. Identify every video-generation endpoint, stop selling commitments that extend past the shutdown, export any metadata your application needs, and design a provider-neutral job interface:
interface VideoJob {
prompt: string;
aspectRatio: "16:9" | "9:16" | "1:1";
durationSeconds: number;
idempotencyKey: string;
}
Put provider-specific fields behind adapters and return an explicit “temporarily unavailable” state if no approved replacement exists. A model abstraction can make later integration easier, but it cannot manufacture a supported Sora replacement.
Which legacy text models shut down on September 28?
gpt-3.5-turbo-instruct, babbage-002, davinci-002, and gpt-3.5-turbo-1106 are scheduled to shut down on September 28. OpenAI recommends gpt-5.6-terra.
This is not always a drop-in replacement. Instruct-style completions, older chat payloads, token usage, and output behavior can differ. Create golden test cases for classification, extraction, rewriting, and refusal behavior. Compare structured-output validity and latency before switching all traffic.
What changes on October 23?
The October 23 wave affects a broad set of older snapshots, including gpt-4-0613, gpt-4-turbo, gpt-4o-2024-05-13, o1-2024-12-17, o3-mini-2025-01-31, and gpt-image-1. It also removes multiple fine-tuned model families.
OpenAI’s replacements vary by model: GPT-5.6 Sol for many high-capability models, GPT-5.6 Terra or Luna for smaller models, gpt-image-2 for image generation, and Sol with reasoning.mode: pro for selected pro models.
Search application code, environment variables, database configuration, prompt-management systems, background workers, and CI fixtures. A model ID can survive in a rarely used queue long after the main application has migrated.
How do you migrate away from OpenAI reusable prompts?
The /v1/prompts API and reusable prompt objects shut down November 30. OpenAI recommends moving prompt content into application code.
Store prompts as versioned application assets rather than inline strings scattered across handlers:
prompts/
support_reply/
v3.system.md
v3.schema.json
evals.jsonl
Record the prompt version with every response. Deploy prompt changes through code review, run a fixed evaluation set, and keep rollback capability. Never put secrets into source-controlled prompt files; inject them from the runtime secret store.
What replaces the OpenAI Evals platform?
Existing OpenAI evals become read-only on October 31, and the dashboard and API shut down November 30. Export definitions, datasets, grader logic, score history, and release thresholds before read-only mode. OpenAI links to Promptfoo as one migration path, but teams can use any system that preserves reproducibility and CI integration.
The critical artifact is not the dashboard. It is the contract connecting a task dataset, prompt version, model version, scoring rule, and release decision. Keep these in a repository or controlled evaluation store so a provider migration does not erase your quality history.
What replaces Agent Builder?
Agent Builder shuts down November 30. OpenAI recommends the Agents SDK or ChatGPT Workspace Agents; ChatKit remains available.
Before moving, document every tool, permission, state transition, approval point, and external side effect. Add idempotency keys to write operations and human approval to destructive actions. A visual workflow can hide important defaults, so migrate behavior—not only boxes on a canvas.
Which GPT Image models end on December 1?
gpt-image-1-mini, gpt-image-1.5, and chatgpt-image-latest shut down December 1. OpenAI recommends gpt-image-2. A separate October 23 deadline applies to gpt-image-1.
Re-run quality, aspect ratio, edit-mask, transparency, content-policy, and latency tests. Image migrations can change output dimensions and file handling even when the prompt is unchanged.
Which GPT-5 and o3 snapshots end on December 11?
The December 11 shutdown includes:
gpt-5-2025-08-07→gpt-5.6-solgpt-5-mini-2025-08-07→gpt-5.6-terragpt-5-nano-2025-08-07→gpt-5.6-lunagpt-5-pro-2025-10-06→gpt-5.6-solwith pro reasoningo3-2025-04-16→gpt-5.6-solo3-pro-2025-06-10→gpt-5.6-solwith pro reasoning
Reasoning settings, output volume, and token prices can change the bill. Use our Astra vs GPT-5.6 Sol agent comparison if the migration is also an opportunity to decide whether Astra is worth its higher per-token price.
Can an OpenAI-compatible API reduce deprecation risk?
It can reduce integration work, but it cannot eliminate model retirement. LLMFly AI is a multi-model AI API platform with an OpenAI-compatible endpoint for selected GPT, Claude, Gemini, and Grok models. Keeping the base URL and model ID in configuration can make a tested model switch faster, and current routes may be priced below official API rates.
Use the live LLMFly AI Model Plaza to verify supported model IDs, pricing, and availability. Do not assume a listed text model replaces provider-specific products such as Sora, Agent Builder, or OpenAI Evals. An API compatibility layer standardizes portable calls; it does not reproduce every hosted tool or product.
For portable text workloads, keep a small abstraction:
const MODEL = process.env.LLM_MODEL;
const BASE_URL = process.env.LLM_BASE_URL;
// Change configuration only after the candidate model passes evals.
const client = new OpenAI({
apiKey: process.env.LLM_API_KEY,
baseURL: BASE_URL,
});Our OpenAI Responses API migration guide covers the broader transition from legacy request patterns.
What is a safe 30-day OpenAI migration plan?
- Days 1–3: inventory models, endpoints, stored prompts, evals, tools, and fine-tuned assets.
- Days 4–7: export configurations and define acceptance tests.
- Days 8–14: implement replacements behind feature flags and provider adapters.
- Days 15–21: shadow traffic, compare cost per accepted result, and inspect failures.
- Days 22–26: canary production traffic with rollback.
- Days 27–30: complete cutover, remove dead credentials, and monitor error rates.
Do not schedule the first production test on the shutdown date. Rate-limit changes, unavailable regions, schema differences, and prompt regressions need time to surface.
Frequently asked questions
Is the OpenAI Assistants API still available?
No. The Assistants API was removed on August 26, 2026. OpenAI recommends the Responses API and Conversations API.
Will GPT-5 stop working in 2026?
Specific older GPT-5 snapshots listed by OpenAI shut down December 11, 2026. This does not mean every current GPT model disappears. Check the exact model ID in logs and configuration.
Does changing a model ID complete an API migration?
Usually not. Reasoning controls, tool calling, output schemas, pricing, latency, and safety behavior can differ. Run regression tests and compare cost per accepted result.
How can I find deprecated model IDs in a codebase?
Search source, environment templates, infrastructure configuration, database records, workflow builders, queues, notebooks, and CI fixtures. Also inspect API usage logs because dynamically stored IDs may not exist in source control.
Bottom line
The 2026 OpenAI deprecation schedule is not one migration. It is a series of product, model, prompt, evaluation, image, and agent deadlines. Start with Sora and September legacy models, then prepare for the October and November waves. A multi-model platform such as LLMFly AI can make portable model switches and price comparisons easier, but provider-specific products still require explicit migration work.
Leave a Reply