Claude Sonnet 4.6 API From China in 2026: Still Available? Pin It or Upgrade to Sonnet 5.5
Sonnet 4.6 is legacy, not retired. What that status means, the request changes that return 400 on Sonnet 5.5, the token-count trap, and how to call 4.6 by API.
1. Is Claude Sonnet 4.6 deprecated? The four lifecycle words
Anthropic uses four terms on its model deprecations page. Active: fully supported and recommended. Legacy: no further updates, might be deprecated in the future. Deprecated: still functional but no longer recommended, with a named replacement and a retirement date. Retired: no longer available, and requests fail. Sonnet 4.6 sits in the second bucket. Its model page says "Legacy. Released February 17, 2026", lists the status as "Active (legacy)", and gives the retirement as "Not sooner than February 17, 2027". The deprecations table shows claude-sonnet-4-6 as Active with no deprecation date yet. So the honest reading is: your integration keeps working today, and the earliest possible retirement is a floor, not a scheduled shutdown. Do not confuse it with Claude Sonnet 4 (claude-sonnet-4-20250514), which was retired on June 15, 2026, or Sonnet 4.5, which the same page lists as deprecated with a November 30, 2026 retirement.
2. How you will hear about a retirement, and how to find your own usage
The deprecations page says Anthropic notifies customers with active deployments by email and in the documentation, with at least 60 days' notice before retiring a publicly released model. Two details matter for planning. First, those dates apply to Anthropic-operated platforms (the Claude API and the platforms Anthropic runs itself); Amazon Bedrock and Google-managed Claude on Google Cloud set their own retirement schedules, so the same model can have different dates depending on where you call it. Second, you can audit which keys still call an old model: in Claude Console open the Usage page, click Export, and read the CSV, which breaks usage down by API key and model. If you call Claude through a relay or gateway instead, ask that provider for the equivalent per-model usage export, because Anthropic's Console only sees traffic sent with your own Anthropic keys. A practical habit: put the model string in one config value, not scattered through code, so a retirement notice becomes a one-line change plus a test run.
3. Good reasons to keep pinning Sonnet 4.6 for now
Anthropic's Model IDs and versioning page explains how IDs, aliases and snapshots relate; check it, then pin the exact ID you tested. Predictable behaviour on code you have already tuned is the main argument for staying. Concrete cases where pinning makes sense: ① you rely on thinking being off by default — on 4.6 a request with no thinking field runs without thinking, while 5.5 runs adaptive thinking unless you send thinking type between_tools; ② your pipeline forces tool calls with tool_choice any or tool, which 5.5 rejects with a 400; ③ you have evals and prompt tuning built around 4.6 output and no time to re-run them this quarter; ④ your cost model is token-sensitive and you have not yet measured the tokenizer change (section 5). One more nuance from the deprecations page: temperature, top_p and top_k are removed in API version 2026-10-21 and return a 400 there for every model, while earlier API versions return that error only for non-default values on Claude 4.7 and later models. If your code still sets temperature on Sonnet 4.6, check which API version header you send before upgrading anything else.
4. Migration checklist: Sonnet 4.6 to Sonnet 5.5
The official Sonnet 5.5 migration guide has a section for callers coming from Sonnet 4.6. Condensed into a checklist you can work through in one branch: ① Remove budget_tokens. thinking type enabled with budget_tokens is deprecated on 4.6 and returns a 400 on 5.5; use adaptive thinking and set effort instead. ② Remove temperature, top_p and top_k. ③ Replace any assistant prefill. Sonnet 4.6 already rejects a prefilled final assistant turn ("This model does not support assistant message prefill"), so if your 4.6 calls work today you probably do not use it, but older code paths written for Sonnet 4.5 may. ④ Rework forced tool choice. ⑤ Expect thinking blocks to look different. On 5.5 thinking text is omitted by default; set display to summarized if you log or show it, which was the default on 4.6. ⑥ Drop old beta headers such as interleaved-thinking-2025-05-14, and move structured output to output_config.format. ⑦ Recount tokens and re-measure image tokens (next section). ⑧ Run your own regression set against claude-sonnet-5-5 before moving production traffic. The guide also suggests letting Claude Code do the mechanical edits with the claude-api migrate command it documents.
5. The cost trap: a lower per-token price is not a lower bill
At the time of writing, Anthropic's model pages list Sonnet 5.5 at a lower per-million-token rate than Sonnet 4.6 for both input and output; check the official pricing page for the current figures before you plan around them. The migration guide then adds two corrections. Text: Sonnet 5.5 uses a newer tokenizer, and the same text produces about 30 percent more tokens than on Sonnet 4.6, depending on content. Images: 5.5 uses a high-resolution tier of up to 2,576 pixels on the long edge, and the guide's example says a 2000×1500 image costs about 2.5 times as many tokens on 5.5 as on 4.6. So the real comparison is price per token multiplied by your token count on each model. A cheap way to measure it: take 50 real requests from your logs, send each to the token counting endpoint with model set to claude-sonnet-4-6 and then claude-sonnet-5-5, and compare totals for input; for output, run a small sample end to end, because adaptive thinking on 5.5 can add output tokens that 4.6 never produced. Only then compare bills. If your traffic is image-heavy, consider resizing to 1,568 pixels on the long edge before sending, which is where 4.6 already stops.
6. Calling Sonnet 4.6 through cocodot (OpenAI-compatible or Anthropic format)
cocodot lists Sonnet 4.6 under the code mcs-5, alongside Sonnet 5 (mcs-6) and Sonnet 5.5 (mcs-7), so you can A/B the two generations with one key. OpenAI-compatible: set base_url to https://cocodot.co/api/ai/v1, use your cocodot key as api_key, and set model to mcs-5; the OpenAI SDK, Cursor's override base URL setting and Cline's OpenAI Compatible provider take these values directly. The official name claude-sonnet-4-6 also resolves to the same model. Anthropic format: if your code uses Anthropic-specific request fields such as thinking or cache_control, call the Anthropic-format endpoint at https://cocodot.co/api/ai instead of translating them through the OpenAI shape, and test the specific fields you rely on with a few requests first. Claude Code: the Claude Code docs describe ANTHROPIC_BASE_URL as the override for routing requests through a gateway, ANTHROPIC_AUTH_TOKEN as the value sent in the Authorization header with a Bearer prefix, and ANTHROPIC_MODEL as the model setting. Set ANTHROPIC_BASE_URL=https://cocodot.co/api/ai (no /v1), ANTHROPIC_AUTH_TOKEN to your cocodot key and ANTHROPIC_MODEL=mcs-5. The docs mark ANTHROPIC_SMALL_FAST_MODEL as deprecated; use ANTHROPIC_DEFAULT_HAIKU_MODEL for the background model instead, for example mch-1.
7. Check you are getting the model you asked for
When you call any model through a third party, "is this really Sonnet 4.6" is a fair question, and a provider's promise does not answer it. Three checks you can run yourself. ① Read the response. The model field in the response echoes the name you requested; that confirms routing intent, not the weights, so treat it as a first filter only. ② Run an independent probe. probe.cocodot.co is an open-source checker: enter any provider's base_url and a temporary key, it runs six probes and returns a score, and the key is discarded after use. Because the code is public you can read what it tests, and you can point it at cocodot or at anyone else. Rotate the temporary key afterwards anyway. ③ Compare against a known baseline. Keep a small fixed set of prompts with answers you have already graded on the official API or a trusted source, re-run it after any provider or model change, and look for drops in the cases that matter to your product. This is the same discipline you need when upgrading from 4.6 to 5.5, so one regression set serves both jobs. For a deeper walk-through, see /hub/ai-relay-model-downgrade-verify.
8. Pricing, caching and payment on cocodot
Listed Claude / GPT / Gemini rates are at or below the official list price; check the pricing page for each model's current rate. The pricing page at cocodot.co/pricing lists every model, including Sonnet 4.6 (mcs-5) and Sonnet 5.5 (mcs-7), so you can repeat the section 5 calculation with our rates as well as Anthropic's. Prompt caching works through cocodot: cache hits at half the input price; cache writes at 1.25x the input price (2x for the portion written to the 1-hour cache), the same structure as the official pricing, and usage reports the cache token counts so you can check them. Be clear-eyed about one point: Anthropic bills cache reads on Sonnet 4.6 at a much smaller fraction of the input price, so for a cache-heavy agent workload, calling Anthropic directly costs less on that part of the bill. Where a relay helps is paying without an accepted card, uncached input and output, and switching between Claude, GPT and Gemini with one key. Billing is per token with no monthly fee. New accounts get a small trial credit after email verification, but it only covers some lower-priced models, so plan a small top-up before testing Sonnet. For the wider choice between going direct and using a relay, see /hub/anthropic-api and /hub/llm-api-vs-official-when.
Sonnet 4.6 vs Sonnet 5.5: what changes for an API caller (source: platform.claude.com model pages and Sonnet 5.5 migration guide, read 2026-10-29)
| Item | Claude Sonnet 4.6 | Claude Sonnet 5.5 | What to do |
|---|---|---|---|
| Lifecycle | Legacy, still active; retirement not sooner than Feb 17, 2027 | Current Sonnet | No rush, but plan tests before any deprecation notice |
| Request with no thinking field | Runs without thinking | Runs with adaptive thinking | Recheck latency and max_tokens on 5.5 |
| thinking.type enabled + budget_tokens | Accepted but deprecated | Returns 400 | Switch to adaptive thinking and the effort parameter |
| Assistant prefill | Returns 400 | Returns 400 | Already broken on 4.6 — use structured output or instructions |
| tool_choice any / tool | Accepted | Returns 400 | Rework forced tool calls before switching |
| Tokens for the same text | Baseline | About 30% more | Recount with the token counting endpoint |
| Minimum cacheable prompt | 1,024 tokens | 512 tokens | Short system prompts may start caching on 5.5 |
| cocodot code | mcs-5 | mcs-7 | Same key and endpoint — change one model string |