Chinese AI Model Deprecation Calendar Q3 2026: DashScope, Anthropic, and OpenAI Sunset Waves

If you run production AI workloads through multiple providers, the next six months look like a minefield. DashScope (Alibaba Cloud Bailian) has four active sunset waves between July 6 and September 8, retiring everything from legacy Qwen snapshots to third-party models like DeepSeek-V3 and GLM-4.7. Anthropic is retiring Claude Opus 4.1 on August 5. OpenAI is pulling GPT-5 and O3 snapshots on December 11, with another batch of legacy models going dark on July 23.

This page is the unified calendar. We built it because no single vendor publishes a cross-provider deprecation view, and if you're routing requests across DashScope, Anthropic, and OpenAI — as many teams do — you need one place to see what's coming and when to act.

We route OpenAI-compatible requests through configured providers and support provider/model routing and fallback where the live product path supports it. We do not claim universal model support, zero downtime, or guaranteed cheapest pricing. Every date and model ID below is sourced from official vendor documentation at retrieval time.

OpenAI-compatible means a provider exposes a chat-completions endpoint whose request and response shape matches the OpenAI API contract closely enough that an unmodified OpenAI SDK call works against it after swapping three values: API key, base URL, and model name. The minimum surface in practice is POST /v1/chat/completions with messages, model, and an OpenAI-shaped streaming response.

30-second answer

• 8 active deprecation waves across 3 major providers between July 2026 and December 2026.
• DashScope is the most aggressive: 4 waves (July 6, 8, 13; September 7, 8) retiring 60+ model IDs including mainstream commercial models like qwen3-max, qwen3-coder-plus, and third-party models like deepseek-v3, glm-4.7 (Model Depreciation, retrieved 2026-06-18).
• Anthropic has one active deprecation: claude-opus-4-1-20250805 retiring August 5, 2026 (Model deprecations, retrieved 2026-06-18).
• OpenAI has two waves: legacy models on July 23, 2026 and GPT-5/O3 snapshots on December 11, 2026 (Deprecations, retrieved 2026-06-18).
• If you use a routing layer with fallback, you can pre-configure replacement models and test before each sunset date hits.

Comprehensive deprecation timeline

The table below lists every upcoming deprecation event sorted by date. Already-retired models are excluded — this is forward-looking only.

Source notes: DashScope dates from Aliyun Model Depreciation page and notices 118344, 118345, 118331, 118332. Anthropic dates from Model deprecations. OpenAI dates from Deprecations. All retrieved 2026-06-18.

DashScope waves: the full breakdown

DashScope's deprecation cadence is the most complex of any provider we track. Four waves overlap in Q3 2026, each targeting different model categories.

July 6 — Voice TTS snapshots (18 days)

Voice-only. Affects qwen-tts-latest, qwen-tts-2025-05-22, qwen-tts-2025-04-10, qwen-tts-realtime-latest, and qwen-tts-realtime-2025-07-15. Replace with cosyvoice-v3-flash. Unless you have TTS in your pipeline, skip this wave.

July 8 — Snapshot models + third-party hosted (20 days)

This is the big one for model diversity. DashScope is sunsetting its hosted copies of DeepSeek, GLM, MiniMax, and Moonshot models alongside 30+ Qwen-family snapshots.

Third-party models affected:

• deepseek-v3, deepseek-v3.1, deepseek-v3.2, deepseek-v3.2-exp, deepseek-r1, deepseek-r1-0528
• deepseek-r1-distill-qwen-7b, deepseek-r1-distill-qwen-14b, deepseek-r1-distill-qwen-32b
• glm-4.7, glm-4.6
• MiniMax-M2.1
• Moonshot-Kimi-K2-Instruct, kimi-k2-thinking

DashScope's recommended replacement for all of these is Qwen-family models (qwen3.7-max, qwen3.7-plus, qwen3.6-flash). But if you chose DeepSeek or GLM for a reason — reasoning depth, coding, or specific benchmark performance — the real migration is routing to the vendor-direct API. See our detailed guide: Migrating Off Bailian-Hosted DeepSeek, GLM, MiniMax, and Moonshot.

Qwen snapshot models affected: qwen3-max-2026-01-23, qwen3-max-2025-09-23, plus 20+ Qwen3 VL, Coder, and open-source snapshots. Full list on the official deprecation page.

July 13 — Mainline models (25 days)

This wave retires the mainline (non-snapshot) versions of the older Qwen generation:

• qwen-turbo and qwen-turbo-realtime — the budget workhorse
• qwen-vl-max and qwen-vl-plus — vision models
• qwq-plus, qvq-max, qvq-plus — reasoning/visual reasoning
• qwen-math-turbo, qwen-coder-turbo, qwen-coder-plus — specialized

Replacements are across the Qwen3.7/3.6 family. This wave has the broadest blast radius for teams using DashScope's "just use the mainline model ID" pattern. See our step-by-step guide: Qwen-Turbo, VL, Coder-Plus 2026-07-13 Sunset Migration Guide.

September 8 — Mainstream commercial (82 days)

The September wave targets current-generation commercial models:

• qwen3.6-max-preview → qwen3.7-max
• qwen3-max-preview → qwen3.7-max
• qwen3-max → qwen3.7-max
• qwen3-vl-flash → qwen3.6-flash
• qwen3-coder-plus → qwen3.7-plus

These are models many teams adopted as "current" within the past 6 months. The replacement models (Qwen3.7 series) bring multimodal upgrades and thinking-mode improvements, but the model ID change is mandatory — there's no auto-migration.

Anthropic lifecycle cadence

Anthropic's deprecation model is the most predictable: at least 60 days between notice and retirement for GA models. The current active deprecation:

• claude-opus-4-1-20250805 — deprecated June 5, 2026. Retiring August 5, 2026 (48 days from now). Recommended replacement: claude-opus-4-8.

The recently retired models (June 15, 2026): claude-opus-4-20250514 and claude-sonnet-4-20250514. If you still have these pinned, your calls are already failing.

For the complete Anthropic deprecation history and migration playbook, see: Anthropic's 60-Day Deprecation Cadence: A Migration Playbook.

OpenAI deprecation waves

OpenAI is running two active deprecation tracks:

July 23 — Legacy model snapshots (35 days)

A broad cleanup of older model variants: gpt-5-chat-latest, gpt-5-codex, gpt-5.1-codex, gpt-5.1-codex-max, o3-deep-research, computer-use-preview, and several audio/realtime snapshots. Replacements are gpt-5.5, gpt-5.4-mini, and gpt-5.5-pro.

December 11 — GPT-5 and O3 snapshots (176 days)

Six specific dated snapshots announced June 11, 2026:

• gpt-5-2025-08-07 → gpt-5.5
• gpt-5-mini-2025-08-07 → gpt-5.4-mini
• gpt-5-nano-2025-08-07 → gpt-5.4-nano
• gpt-5-pro-2025-10-06 → gpt-5.5-pro
• o3-2025-04-16 → gpt-5.5
• o3-pro-2025-06-10 → gpt-5.5-pro

OpenAI's minimum notice for GA models is 6 months; preview models may get as little as 2 weeks. The O3 models are particularly noteworthy because their reasoning architecture is being folded into the GPT-5.5 line — it's not a 1:1 replacement in behavior, even if the API surface is compatible.

Source: OpenAI Deprecations, retrieved 2026-06-18.

Decision tree: "I use model X — what do I do?"

Use this flowchart to determine your action:

1. Is your model on the timeline above?

• No → You're fine for now. Bookmark this page and check back monthly.
• Yes → Continue.

2. How many days until sunset?

• ≤ 30 days (July 6, 8, 13 waves) → Act now. Test replacements this week, deploy next week.
• 30–90 days (July 23, Aug 5, Sep 8 waves) → Plan now, test next month. Put it on the sprint board.

• 90 days (Dec waves) → Awareness only. No action needed yet; re-check in September.

3. Is the replacement a same-vendor upgrade or a vendor switch?

• Same-vendor upgrade (e.g., qwen3-max → qwen3.7-max) → Test the replacement model with your eval suite. Check for behavior differences in thinking mode, vision inputs, or pricing. Deploy when satisfied.
• Vendor switch (e.g., deepseek-v3 on DashScope → api.deepseek.com direct) → You need to add the new provider endpoint and authentication. A routing layer simplifies this — configure both endpoints, run shadow traffic, then cut over.

4. Do you have fallback configured?

• Yes → Add the replacement model to your fallback chain now. When the old model goes dark, traffic automatically routes to the replacement.
• No → Consider adding a fallback layer. See our guide: LLM API Fallback Routing Guide.

FAQ

How much notice does DashScope give before retiring a model?

DashScope provides 30 days' notice for snapshot models (those with date identifiers like qwen3-max-2026-01-23) and 3 months' notice for mainline models (core versions like qwen-turbo or qwen3-max). Notice is delivered via SMS, email, in-console messages, and official website announcements. Source: Model Depreciation, retrieved 2026-06-18.

What happens if I keep calling a retired model?

On DashScope, the model inference service stops completely on the retirement date. API calls return errors, and any applications or services using the model will fail to produce results. On Anthropic and OpenAI, retired models also return API errors. There is no grace period after the stated retirement date on any provider.

Can I keep using DeepSeek or GLM models after DashScope removes them?

Yes — DashScope is only removing its hosted copies. The vendor-direct APIs (api.deepseek.com, open.bigmodel.cn, platform.kimi.ai) are unaffected. You need to switch your endpoint URL and API key. See: Migrating Off Bailian-Hosted DeepSeek, GLM, MiniMax, and Moonshot.

How does the September 8 wave differ from the July waves?

The July waves retire snapshots and legacy mainline models. The September 8 wave retires current-generation commercial models (qwen3-max, qwen3-coder-plus, qwen3-vl-flash) — these are models many teams adopted in late 2025 or early 2026. The replacements (Qwen3.7 series) bring architectural improvements including multimodal support in qwen3.7-max, but the switch is not automatic.

Does TheRouter automatically handle these model transitions?

TheRouter routes OpenAI-compatible requests through configured providers and supports provider/model routing and fallback where the live product path supports it. You can pre-configure replacement model IDs in your routing config and set up fallback chains, so when a deprecated model starts returning errors, traffic routes to the replacement. But you need to configure this — it's not automatic out of the box. See: OpenAI to TheRouter Migration Guide.

What's the cheapest migration path for teams on DashScope?

For teams fully committed to Qwen models: follow DashScope's recommended replacements — they're same-vendor, same-endpoint upgrades. For teams that chose third-party models for specific capabilities: route to the vendor-direct API. In many cases, vendor-direct pricing is competitive or cheaper than DashScope hosted pricing, especially for DeepSeek models which offer aggressive pricing on their own platform. See: DeepSeek API Complete Guide and Aliyun Bailian API Guide.

Routing and fallback during sunset transitions

The safest migration pattern during any deprecation wave:

from openai import OpenAI

# Configure primary and fallback through a routing layer
client = OpenAI(
    base_url="https://api.therouter.ai/v1",
    api_key="<THEROUTER_API_KEY>",
)

# Request with a model that's being deprecated
response = client.chat.completions.create(
    model="qwen3-max",       # Retiring Sep 8
    messages=[{"role": "user", "content": "Explain transformer attention"}],
)

# With routing + fallback configured:
# 1. Before Sep 8: routes to qwen3-max on DashScope
# 2. After Sep 8: automatically falls back to qwen3.7-max
# 3. No code change needed — just routing config

The key is configuring the fallback before the sunset date, not after your production starts throwing errors at 3 AM. For the full fallback routing setup, see: LLM API Fallback Routing Guide.

What's next

We update this calendar as providers announce new deprecations. The next expected events:

• DashScope tends to announce deprecation waves quarterly. Expect Q4 2026 announcements around September–October.
• Anthropic follows a roughly 60-day cadence tied to new model releases. When Opus 4.9 or Opus 5 ships, expect Opus 4.5/4.6 deprecation notices.
• OpenAI has been accelerating deprecation cadence as model versions proliferate. The December 2026 wave is likely not the last for GPT-5 era snapshots.

Bookmark this page. We'll keep it current.

---

All deprecation dates sourced from official vendor documentation. Retrieved 2026-06-18. Dates are subject to change by the respective providers — always verify against the official source before taking action.