AI API deprecation calendar 2026: OpenAI and Google migration deadlines to watch
- →Google currently lists October 22, 2026 for the shutdown of several Veo 3.1 preview model IDs.
- →OpenAI has multiple legacy model snapshots scheduled to shut down October 23, 2026; Evals becomes read-only October 31.
- →OpenAI Evals, Agent Builder, and reusable prompt objects are scheduled to shut down November 30, 2026.
- →The safest migration is not a string replacement: inventory usage, test the replacement on real fixtures, validate tools and permissions, then remove the deprecated dependency.
AI platform deprecations have become an engineering operations problem, not just release-note housekeeping. A model ID can live in application code, environment variables, workflow exports, prompt configuration, CI jobs, notebooks, and vendor integrations. Missing one reference can turn a published shutdown date into a production incident.
The near-term AI deprecation calendar
| Provider | Model / feature | Milestone | Current migration direction |
|---|---|---|---|
| Google Gemini API | veo-3.1-lite-generate-preview | Shutdown listed: Oct 22, 2026 | gemini-omni-1.1-flash |
| Google Gemini API | veo-3.1-generate-preview | Shutdown listed: Oct 22, 2026 | gemini-omni-1.1-flash |
| Google Gemini API | veo-3.1-fast-generate-preview | Shutdown listed: Oct 22, 2026 | gemini-omni-1.1-flash |
| OpenAI | Legacy model snapshots | Shutdown: Oct 23, 2026 | Move to the current replacement documented for each model |
| OpenAI | Evals platform | Read-only: Oct 31, 2026 | Export/migrate evaluation workflows before write access disappears |
| OpenAI | Evals platform | Shutdown: Nov 30, 2026 | OpenAI documents Promptfoo as one migration path |
| OpenAI | Agent Builder | Shutdown: Nov 30, 2026 | Agents SDK or ChatGPT Workspace Agents, with parity testing |
| OpenAI | Reusable prompt objects / v1/prompts | Shutdown: Nov 30, 2026 | Move production prompts into application code |
| OpenAI | Legacy text-to-speech models | Shutdown: Jan 6, 2027 | gpt-realtime-2.1-mini |
| OpenAI | gpt-5.3-codex, gpt-5.4-nano, gpt-5.1 | Shutdown: Apr 1, 2027 | Current GPT-6 replacements listed by OpenAI |
Deprecation dates change the priority of engineering work. Always re-check the provider documentation before making a production migration decision; do not rely on an old spreadsheet or a copied blog post.
1. Inventory where the deprecated dependency actually lives
Searching only application source code is not enough. Model IDs and feature references often sit in environment variables, workflow JSON, infrastructure configuration, notebooks, CI/CD variables, serverless functions, SDK wrappers, and vendor-specific settings. Start by building an inventory of every exact model ID, endpoint, or platform feature in use.
- Search exact model IDs and aliases, not only provider names.
- Inspect .env templates and deployment configuration without exposing secrets.
- Scan workflow exports and low-code automation definitions.
- Check scheduled jobs, notebooks, tests, examples, and fallback code paths.
- Separate a documented alias from a pinned snapshot because their retirement behavior may differ.
2. Rank by deadline and blast radius
A dependency used in a dormant demo is not the same risk as one inside checkout, customer support, document processing, or a production agent with tool access. Score each hit by days to shutdown, production traffic, business criticality, replacement uncertainty, and rollback difficulty.
| Status | Meaning | Next action |
|---|---|---|
| READY | Replacement is tested and deployment path is known | Schedule migration and rollback window |
| AT RISK | Dependency exists and replacement is identified, but parity is not proven | Run fixtures, evals, and integration tests |
| BLOCKED | Dependency exists but owner, replacement, access, or migration path is unresolved | Escalate immediately and assign an owner |
3. Prove behavior, not just API compatibility
A migration can compile and still be wrong. New models may change tool selection, latency, cost, output shape, refusal behavior, context handling, or structured-output reliability. Agent migrations add another layer: permissions, authentication, control flow, and human approval points must still behave correctly.
- Capture representative production-safe fixtures.
- Run the old and replacement path side by side where possible.
- Compare task success, structured outputs, latency, cost, and failure modes.
- For agents, validate tool calls, permissions, approval boundaries, and rollback behavior.
- Use a canary or staged rollout before removing the old path.
4. Treat prompt and eval migrations as code migrations
OpenAI is moving reusable prompt objects toward prompts stored in application code. That is useful operationally: code-managed prompts can use typed inputs, code review, tests, fixtures, and normal deployment controls. The same principle applies to eval migration: preserve the dataset, grader intent, thresholds, and pass/fail meaning, not merely the file format.
5. Keep a source-backed deprecation registry
The recurring product opportunity is larger than one OpenAI shutdown. A useful deprecation registry stores the provider, exact feature or model ID, announcement date, shutdown date, replacement, source URL, and the date the rule was last verified. A repository scan can then turn that registry into concrete migration work.