benchr
Deprecation Tracker ·
AI model deprecation tracker
Announced retirements, confirmed sunset dates, and provider-listed migration targets — with compatibility testing required before production. Migrate before the sunset date to avoid broken workloads.
Next retirement
Gemini Robotics ER 1.6 · Kimi K2.5 + Moonshot V1
Google · Moonshot AI · August 31, 2026
Full migration guides, with price & benchmark deltas →
No models match these filters.
Announced & confirmed sunsets
| Model | Provider | Status | Sunset date | Migrate to | Source |
|---|---|---|---|---|---|
| Assistants API | OpenAI | Retired | Aug 26, 2026 | Responses API + Conversations API — OpenAI gave a full year of notice on August 26, 2025 and the date passed on August 26, 2026 — OpenAI now files the platform under past deprecations. Threads become conversations, runs become responses; budget real testing time, because the request shape changes, not just the model ID. | developers.openai.com/api/docs/deprecations |
| Imagen 4 API models (standard, Ultra, Fast) | Earliest date Aug 17 | Aug 17, 2026 (earliest — passed) | Gemini 3.1 Flash Image — Google's listed replacement for imagen-4.0-generate-001, imagen-4.0-ultra-generate-001, and imagen-4.0-fast-generate-001. The published date is the earliest possible shutdown; Google says it will communicate the exact date in advance. |
ai.google.dev/gemini-api/docs/deprecations | |
| Gemini Robotics ER 1.6 Preview | Deprecated | Aug 31, 2026 | Gemini Robotics ER 2 Preview — gemini-robotics-er-2-preview replaces gemini-robotics-er-1.6-preview. Announced with the ER 2 public preview on July 30, 2026. Preview endpoints get short notice by design. |
ai.google.dev/gemini-api/docs/deprecations | |
| Kimi K2.5 and Moonshot V1 series | Moonshot AI | Retired | Aug 31, 2026 | Kimi K3 — kimi-k2.5 and all six moonshot-v1 text/vision endpoints are already closed to newly registered users. Moonshot's docs mark the series retired on August 31, 2026, and calls now return a 404. |
platform.kimi.ai/docs/models |
| Videos API · Sora 2 · Sora 2 Pro | OpenAI | Deprecated | Sep 24, 2026 | No replacement listed — OpenAI's table shows a dash. If you generate video through sora-2 or sora-2-pro, you're choosing a new vendor, not swapping an identifier. |
developers.openai.com/api/docs/deprecations |
| Legacy audio, realtime & transcription models | OpenAI | Deprecated | Jan 20, 2027 | GPT Realtime 2.1 / GPT Audio 1.5 — announced July 20, 2026 and covering nine IDs: gpt-realtime, gpt-audio, gpt-4o-realtime, gpt-4o-audio, their mini variants, and gpt-4o-mini-transcribe-2025-03-20. Realtime IDs move to gpt-realtime-2.1, audio IDs to gpt-audio-1.5, transcription to the 2025-12-15 snapshot. |
developers.openai.com/api/docs/deprecations |
| Imagen 4.0 (generate · ultra · fast) | Retired | Aug 17, 2026 | Gemini 3.1 Flash Image — all three Imagen 4.0 generation IDs shut down August 17, 2026. Google folds image generation into the Gemini line rather than shipping an Imagen 5. | ai.google.dev/gemini-api/docs/deprecations | |
| Prompt tools API + Workbench | Anthropic | Retired | Aug 17, 2026 | No successor — the experimental generate, improve, and templatize prompt endpoints retired with the Workbench on August 17, 2026. Prompt tooling moves into your own codebase. No Claude model changed status alongside it. | platform.claude.com/docs |
| GPT-5.2 / GPT-5.3 chat-latest | OpenAI | Retired | Aug 10, 2026 | GPT-5.6 Sol — OpenAI's official replacement, and it got 20% cheaper on input the same month. The deadline has passed and OpenAI now files the record under past deprecations, so gpt-5.2-chat-latest and gpt-5.3-chat-latest no longer answer. |
developers.openai.com/api/docs/deprecations |
| Gemini Embedding 2 Preview | Retired | Aug 10, 2026 | Gemini Embedding 2 — swap embedding-2-preview for gemini-embedding-2. The August 10 date has passed and Google greys out shut-down models on its table. Re-embed anything still keyed to the preview: vectors from two model versions don't share a space. |
ai.google.dev/gemini-api/docs/deprecations | |
| Claude Sonnet 4 | Anthropic | Retired | Jun 15, 2026 | Claude Sonnet 4.6 — same $3/$15 pricing, 1M context. Swap model id before the date; calls fail after retirement. Migration guide. | platform.claude.com/docs |
| Claude Opus 4 | Anthropic | Retired | Jun 15, 2026 | Claude Opus 4.8 — Anthropic's listed replacement, with a lower $5/$25 rate than the retired $15/$75 snapshot. Validate output and prompts before swapping the model ID. Migration guide. | platform.claude.com/docs |
| Claude Opus 4.1 | Anthropic | Retired | Aug 5, 2026 | Claude Opus 4.8 — Anthropic's listed migration target with lower snapshot pricing; capability differences depend on the workload. Migration guide. | platform.claude.com/docs |
| GPT-5.3 Instant (ChatGPT) | OpenAI | Phasing out | ~Aug 2026 | GPT-5.5 Instant — already the ChatGPT default since May 5; paid users keep 5.3 Instant roughly three months from that date. | help.openai.com release notes |
| Gemini 3 Pro | Sunset | Mar 9, 2026 | Gemini 3.1 Pro — provider-listed successor in the same pricing tier. Update the ID only after testing request fields, prompts, output parsing, safety behavior, and task quality. (Vertex AI listed the prior preview discontinued by Mar 26.) | ai.google.dev/deprecation | |
| Claude Mythos Preview | Anthropic | Deprecated | Not published | Claude Mythos 5 inside Project Glasswing. General developers: Claude Fable 5 is the public Mythos-class model. | platform.claude.com/docs |
| deepseek-chat / deepseek-reasoner | DeepSeek | Retired | Jul 24, 2026 | DeepSeek V4-Flash / V4-Pro — alias rename only, no pricing or capability change. The legacy names are no longer available. | api-docs.deepseek.com |
| Claude Opus 4.7 fast mode | Anthropic | Retired | Jul 24, 2026 | Claude Opus 4.8 fast mode or Opus 4.7 standard mode. Fast mode removal only; base Opus 4.7 remains available. | platform.claude.com/docs |
| Gemini 3 Pro Image Preview | Retired | Jun 25, 2026 | Gemini 3 Pro Image (stable) — the stable release replaced the preview. Test ID availability, image inputs, response fields, and output quality before production cutover. | ai.google.dev/deprecation | |
| Codex models, deep-research, computer-use-preview | OpenAI | Retired | Jul 23, 2026 | GPT-5.5 (Codex), GPT-5.5 Pro (deep-research), GPT-5.4 mini (computer use, codex-mini). Wave breakdown. | developers.openai.com/api/docs/deprecations |
| Gemini 2.5 Pro + 2.5 Flash | Deprecated | Oct 16, 2026 | Gemini 3.1 Pro / Gemini 3.5 Flash — both raise list prices; Flash input jumps 5×. Migration guide. | ai.google.dev/deprecation | |
| GPT-4o (2024-05-13) | OpenAI | Deprecated | Oct 23, 2026 | GPT-5.5 is OpenAI's pick; GPT-5 is the cost match for most 4o traffic. Migration guide. | developers.openai.com/api/docs/deprecations |
| GPT-4, GPT-4 Turbo, GPT-3.5, GPT-4.1 nano, GPT Image 1, o1/o3-mini/o4-mini | OpenAI | Deprecated | Oct 23, 2026 | GPT-5.5 / GPT-5.4 mini per OpenAI. Eleven model IDs go the same day. Full wave breakdown. | developers.openai.com/api/docs/deprecations |
| Prompts API · Evals · Agent Builder | OpenAI | Deprecated | Nov 30, 2026 | Platform products, not models (announced June 3, 2026). Agent Builder's official path is the Agents SDK; prompts move into application code. | developers.openai.com/api/docs/deprecations |
| Older GPT-5 / o3 snapshots | OpenAI | Deprecated | Dec 11, 2026 | Dated snapshots only (gpt-5-2025-08-07, o3-2025-04-16, …) — the floating gpt-5 / gpt-5-mini / gpt-5-nano aliases stay live. Per-id replacements: GPT-5.5 / GPT-5.4 mini / GPT-5.4 nano / GPT-5.5 Pro. Announced June 11, 2026. | developers.openai.com/api/docs/deprecations |
| Claude 3 Opus | Anthropic | Retired | Jan 5, 2026 | Claude Opus 4.8 — Anthropic's listed replacement. Calls to claude-3-opus now fail. | platform.claude.com/docs |
| Claude 3.5 Haiku | Anthropic | Retired | Feb 19, 2026 | Claude Haiku 4.5 — same pricing tier, better capability. Retired alongside Claude Sonnet 3.7. | platform.claude.com/docs |
| Gemini 1.0 Pro | Sunset | Apr 9, 2026 | Gemini 3.1 Pro or Gemini 3.5 Flash depending on quality/cost needs. | ai.google.dev/deprecation |
Need the full story on any of these? The deprecations hub keeps the permanent record — every retirement with its price delta, benchmark delta, and migration guide, plus an RSS feed for new announcements. Calls to a retired ID already failing? The API error database covers the 404s and the fixes.
Active models (no announced sunset)
These models have no announced end-of-life in the record as of August 24, 2026. Anthropic is the only provider that publishes a tentative retirement floor per model, so its rows carry a "not sooner than" date; the others announce a sunset only when they schedule one. Model generations move fast, so follow the RSS update feed below for new provider announcements.
| Model | Provider | Status | Notes |
|---|---|---|---|
| Claude Opus 5 | Anthropic | Active | Released July 24, 2026. Anthropic's floor: not sooner than July 24, 2027. |
| Claude Sonnet 5 | Anthropic | Active | Floor: not sooner than June 30, 2027. |
| Claude Fable 5 | Anthropic | Active | Released June 9, 2026. Floor: not sooner than June 9, 2027. |
| Claude Opus 4.8 | Anthropic | Active | Released May 28, 2026. Active high-end Opus tier in this snapshot. |
| Claude Opus 4.7 | Anthropic | Active | Superseded by 4.8 but still supported. Floor: April 16, 2027. Its fast mode retired July 24, 2026. |
| Claude Opus 4.6 | Anthropic | Active | Released February 5, 2026. Floor: not sooner than February 5, 2027. |
| Claude Opus 4.5 | Anthropic | Active | Floor: not sooner than November 24, 2026 — the nearest Claude floor on the table. |
| Claude Sonnet 4.6 | Anthropic | Active | Daily-driver Sonnet tier. Floor: not sooner than February 17, 2027. |
| Claude Sonnet 4.5 | Anthropic | Active | Floor: not sooner than September 29, 2026. |
| Claude Haiku 4.5 | Anthropic | Active | Small/fast tier. Floor: not sooner than October 15, 2026. |
| GPT-5.6 Sol / Terra / Luna | OpenAI | Active | GA July 9, 2026. Sol's price fell to $4 in / $20 out per 1M on August 21, promotional at least through November 21, 2026. |
| GPT-5.5 | OpenAI | Active | Released April 2026. Superseded as flagship by the GPT-5.6 series in July. |
| GPT-5 | OpenAI | Active | Launched Aug 2025. Main production model. |
| GPT-5 Mini | OpenAI | Active | Small/cheap tier. |
| Gemini 3.7 Flash | Active | GA August 13, 2026. Introductory $0.75 in / $3.75 out per 1M through December 31, 2026. | |
| Gemini 3.6 Flash | Active | GA July 21, 2026. Now priced identically to 3.7 Flash. | |
| Gemini 3.5 Flash | Active | GA since May 19, 2026. | |
| Gemini 3.5 Flash-Lite | Active | GA July 21, 2026. The cheapest Gemini subagent tier. | |
| Gemini 3.1 Pro | Active | Preview. Replacement for Gemini 3 Pro. | |
| Grok 4.6 | xAI | Active | Announced August 12, 2026. 500K context, $2 in / $6 out per 1M below 200K tokens. |
| Grok 4.5 | xAI | Active | Released July 8, 2026. Same headline rate as 4.6, cheaper cached input. |
| Grok 4.3 | xAI | Active | The 1M-context Grok. Superseded as flagship by 4.5, then 4.6. |
| DeepSeek V4-Pro / Flash | DeepSeek | Active | MIT license. Self-hosted or API. |
| Kimi K2.6 | Moonshot AI | Active | Modified MIT. |
| Mistral Large 3 | Mistral | Active | Apache 2.0. Dec 2025. |
| Llama 4 Scout / Maverick | Meta | Active | Community license. Self-host. |
| Qwen 3.6 | Alibaba | Active | Apache 2.0. |
| Phi-4 | Microsoft | Active | MIT. Edge/local use. |
How to migrate safely
A model-ID change may be the first code edit, but that change does not establish compatibility. Request schemas, supported tools, context limits, safety behavior, output shape, latency, cost, and task quality can differ. Use this compatibility process before cutover:
- Read the provider's official migration guide — they almost always include context-window and capability notes.
- Run your top 10 prompt templates against both the old and new model. Compare outputs.
- Check for prompt-length changes — newer models sometimes handle system prompts differently.
- Update your model ID in staging first; test for at least 48 hours before production.
- Monitor cost — newer models often have different pricing even at the "same tier."
Need to quickly compare the old and new model? Use the side-by-side compare tool.
Frequently asked questions
What is model deprecation?
When a provider announces that a model version will stop accepting API calls after a specific date. Usually they give 30–90 days notice and suggest a successor model. After the sunset date, calls to that model ID will fail or redirect.
How much notice do providers give?
It varies. Anthropic and OpenAI typically give 30–90 days notice. Google has used 30-day windows. Some smaller providers give less. Follow the RSS update feed below for new entries.
Is Gemini 3 Pro still accessible?
No — Gemini 3 Pro Preview was shut down March 9, 2026 on the Gemini API (Vertex AI listed it discontinued by March 26). Migrate to Gemini 3.1 Pro (same pricing, better reasoning) or Gemini 3.5 Flash (faster, cheaper, with a free tier).
Changelog
- — Removed automatic-compatibility claims, added explicit compatibility testing, and refreshed the visible tracker date.