Claude Sonnet 4.6 retirement date: Opus 4.6 and Sonnet 4.6 in February 2027

Summary. Anthropic's model deprecations page lists claude-opus-4-6 with a tentative retirement date of not sooner than 5 February 2027 and claude-sonnet-4-6 not sooner than 17 February 2027. On 1 November 2026 both are still marked Active there, and Anthropic promises at least 60 days' notice, so a formal notice would have to arrive by 7 and 19 December 2026 to hold those dates. Microsoft Foundry's retirement schedule is firmer and earlier: 2 February 2027 for claude-opus-4-6 and 10 February 2027 for claude-sonnet-4-6, with no replacement listed. Amazon Bedrock's model cards show the same 5 and 17 February floors, but neither model has entered Legacy, and the older-model policy requires at least six months there before end of life, so Bedrock access should last beyond February. The likely successors cost less per token: Claude Sonnet 5.5 at $2 input and $10 output per million tokens against $3 and $15, and Claude Opus 5.5 at $4 and $20 against $5 and $25. The catch is behaviour: thinking runs by default, temperature and thinking budgets return 400 errors, forced tool choice is rejected, and the same text produces about 30 percent more tokens on Sonnet and up to 35 percent more on Opus. Plan around the earliest date on any platform you use.

About this article. This guide is for engineering leads and developers running Claude Opus 4.6 or Sonnet 4.6 on the Claude API, Bedrock, Foundry or Google Cloud. Dates, prices and API behaviour were checked on 1 November 2026 against Anthropic, AWS and Microsoft documentation, and we ran the code below through Anthropic's Python SDK against a local mock server. Vendors change schedules and prices often, so re-check before you commit; no vendor paid for inclusion. Researched with AI assistance and reviewed by our editorial team.

Three platforms, three answers for February 2027

The same two models carry different dates depending on where you call them.

Anthropic's floor dates. Anthropic's Deprecated state means a replacement is named and a retirement date assigned. Neither 4.6 model has reached it, though its models overview already files both under legacy models, meaning no longer updated and possibly deprecated later. Not sooner than is a floor; the actual date can be later. The dates apply to Anthropic-operated platforms, including Microsoft Foundry. Bedrock and the Google-managed offering set their own.

Microsoft's firm dates. The Foundry schedule, last updated on 23 September 2026, lists claude-opus-4-6 (version 1) retiring on 2 February 2027 and claude-sonnet-4-6 (version 1) on 10 February 2027, both GA, neither with a named replacement. Microsoft's policy says GA models from Anthropic follow a 12-month lifecycle instead of the standard 18 months, retired models return 410 Gone, and retirement dates cannot be extended. It names replacements about 90 to 120 days before retirement and sends notices at least 60 days before, which points to names by about 4 and 12 November 2026 and notices by 4 and 12 December.

Bedrock's later horizon. Both 4.6 models follow Bedrock's legacy policy. Their cards show launches on 5 and 17 February 2026, floors one year later and no EOL date, and neither appears among models with a scheduled EOL. A model must spend at least six months in Legacy, so a notice issued on 1 November would put EOL no earlier than May 2027. That is our inference, not a promise: the same AWS table shows older Claude models in ap-south-1 with only 34 days between Legacy and EOL, so check the Region you use. The precedent is Claude Sonnet 4.5: Anthropic retires it on the Claude API on 30 November 2026, while Bedrock lists EOL on 8 April 2027.

Why the dates disagree, and which one to plan around

The mismatch is not unique to the 4.6 models: Foundry's dates are earlier for some models and later for others.

Anthropic says its dates apply to Foundry, and its overview says Foundry follows the Claude API schedule, yet Microsoft publishes different ones, so at least one page is out of step. Floors can also slip later: Haiku 4.5's passed 17 days ago and it is still Active.

Our planning rule: finish cut-over two weeks before the earliest date on any platform you call. For a team on the Claude API and Foundry, that means Opus 4.6 traffic at zero by 19 January 2027 and Sonnet 4.6 by 27 January 2027. For other vendors, see our comparison of AI model retirement policies.

To find your exposure, export the CSV from the Claude Console Usage page, which breaks usage down by API key and model. The Models API reports lifecycle and retires_at for each model; retires_at stays null while a model is active, so a nightly job can alert on any change. On Bedrock, read modelLifecycle from GetFoundationModel in each Region. On Foundry, set an Azure Service Health alert for Azure OpenAI Service.

Replacement mapping and current prices

Anthropic has not named replacements for the 4.6 models, because they are not deprecated. For Sonnet the pattern is clear: Sonnet 4.5, Sonnet 4 and Sonnet 3.7 all list claude-sonnet-5-5 as the replacement. For Opus it is mixed: the deprecations page lists claude-opus-4-8 as the replacement for Opus 4.1, Opus 4 and Opus 3, while the models overview says to start with Opus 5.5 for most workloads. Prices below are per million tokens from Anthropic's pricing pages, with five-minute cache writes. Floors are Anthropic's.

All four current models have a 1M-token context window and 128K max output, and Batch halves their prices. Opus 5.5, released on 22 September 2026, offers fast mode at 2x standard pricing. Bedrock's cards give Opus 5.5 and Sonnet 5.5 the same floors, with a six-month Legacy period. Foundry's schedule has no row yet for claude-opus-5-5 or claude-sonnet-5-5, only claude-opus-5-5-fast at 15 January 2027, so confirm the target's date there before you commit.

Opus 4.8 and Sonnet 5 accept settings the 5.5 models reject, such as thinking.type set to disabled and forced tool choice. They buy about four to seven months, then you migrate again, and fast mode on Opus 4.8 ends on 30 November 2026. Choose them only if a hard dependency blocks the 5.5 move before February.

What breaks when you move from 4.6 to the 5.5 models

Anthropic's migration guides list these changes for code coming from the 4.6 generation.

Thinking now runs by default

On Sonnet 4.6, a request with no thinking field ran without thinking. On Sonnet 5.5 the same request runs with adaptive thinking, and on Opus 5.5 thinking is always on and cannot be disabled. Thinking tokens bill as output and count against max_tokens. On Sonnet 5.5, thinking.type set to disabled returns a 400 error; the lowest setting is between_tools, accepted only at low, medium or high effort. Responses can now start with thinking blocks, so code that reads content[0].text breaks. Read blocks by type, and pass thinking blocks back unchanged in tool loops.

Settings that now return 400 errors

• Thinking budgets. thinking.type set to enabled with budget_tokens was deprecated but accepted on 4.6. Both 5.5 models reject it. Use output_config.effort, and run evals at two or three effort levels, because there is no fixed mapping from a budget to an effort level.

• Sampling parameters. Remove temperature, top_p and top_k. API version 2026-10-21 removes them for every model, and earlier versions reject non-default values on Claude 4.7 and later.

• Forced tool choice. A tool_choice of type any or tool is rejected. Use auto with strict tools and say in the prompt when to call them. A request can carry at most 20 strict tools.

Defaults that change silently

• Effort. Opus 5.5 defaults to medium effort, where Opus 4.6 defaulted to high. Sonnet 5.5 defaults to high on the Claude API. Set effort explicitly on every call.

• Thinking display. Thinking text is omitted by default on the 5.5 models, where both 4.6 models returned summaries. Set display to summarized if your interface shows reasoning.

• Text between tool calls. Longer notes between tool calls now arrive as thinking blocks that are empty by default, so a progress feed can go quiet without any error.

Token counts, images and caching

Sonnet 5.5 produces about 30 percent more tokens than Sonnet 4.6 for the same text. Opus 5.5 uses the tokenizer introduced with Opus 4.7, which Anthropic says may use roughly 1x to 1.35x as many tokens as earlier Opus models. Images also cost more: a 2000 x 1500 image costs about 2.5 times as many tokens on Sonnet 5.5. The minimum cacheable prompt drops from 1,024 tokens to 512.

Conversation and safety changes

Thinking blocks on the 5.5 models are signed over the conversation before them. On accounts created on or after 31 August 2026, replaying a block after editing earlier history returns a 400 error unless you set thinking.block_binding.prefix_mismatch_behavior, so keep conversations append-only. Real-time cyber safeguards are new relative to both 4.6 models, refusals arrive with stop_reason value refusal and a category such as cyber or bio, and Opus 5.5 does not support Priority Tier.

A minimal before-and-after for a typical Sonnet 4.6 call:

            # Before: Claude Sonnet 4.6 (anthropic Python SDK 0.x)
client.messages.create(
    model="claude-sonnet-4-6",
    max_tokens=4000,
    temperature=0.2,
    tool_choice={"type": "tool", "name": "extract_invoice"},
    tools=tools,
    messages=messages,
)

# After: Claude Sonnet 5.5
resp = client.messages.create(
    model="claude-sonnet-5-5",
    max_tokens=8000,                      # covers thinking plus text
    output_config={"effort": "medium"},
    tool_choice={"type": "auto"},
    tools=[{**t, "strict": True} for t in tools],
    messages=messages,                    # prompt says when to call the tool
)
text = "".join(b.text for b in resp.content if b.type == "text")
          

A 4.6 call with a thinking budget cannot also force a tool, so budget-based calls need only the budget replaced with an effort level. Against our mock server, Python SDK 0.125.0 accepted the first call and 2.3.0 the second; 1.17.0 and 2.3.0 raised a TypeError on temperature.

Anthropic offers /claude-api migrate this project to claude-sonnet-5-5 in Claude Code for the mechanical edits; it does not tell you whether quality held.

The Bedrock trap: API version 2026-10-21

On 21 October 2026 Anthropic released SDKs that send API version 2026-10-21 by default, including Python 2.0.0 and TypeScript 1.0.0. Its legacy Bedrock guide warns that this default also applies through Bedrock, and that the ARN-versioned models in its table, which include anthropic.claude-opus-4-6-v1 and anthropic.claude-sonnet-4-6, do not support that version there. Requests can return a 400 error. Pin the SDK you run today wherever you install it, which for Python means staying below 2.0.0. Boto3 code that sets anthropic_version in the request body is not affected.

Off Bedrock, the same upgrade turns any leftover temperature into a 400 error on the 4.6 models too, because the new version removes it for every model.

Once a Legacy notice lands, existing customers can lose access after 15 days of inactivity, which catches quarterly batch jobs, and after at least three months the model can move to public extended access at higher prices. Opus 5.5 is on Bedrock as anthropic.claude-opus-5-5, served through the Messages-API endpoint at /anthropic/v1/messages.

Decision table: which model to move to

For model choice by task type, our 2027 coding model comparison covers Opus against GPT and Gemini options, and our Gemini 3.8 Flash, Claude Haiku 5.5 and GPT-6 Luna comparison covers the budget tier.

Worked cost examples

Illustrative arithmetic from list prices; eCorpIT did not benchmark these models. Token volumes are per month, in millions.

Scenario A: a Sonnet 4.6 coding agent. 2,000M input tokens, of which 1,600M are cache reads, 100M are cache writes and 300M are uncached, plus 150M output tokens.

The tokenizer takes half the saving and default thinking can take the rest. Measure output tokens per task at each effort level.

Scenario B: an Opus 4.6 analysis workload. 300M cache reads, 20M cache writes, 80M uncached input and 40M output a month. Opus 4.6 costs $150 + $125 + $400 + $1,000 = $1,675. Opus 5.5 with identical token counts costs $60 + $100 + $320 + $800 = $1,280, a 24% saving. At 1.2x tokens it is $1,536; at the 1.35x upper bound it is $1,728, 3% more than today. If Sonnet 5.5 passes your evals, the bill at 1.3x tokens is about $832. Rerun these numbers with our LLM token counter.

India, EU and US data residency notes

India. AWS's cards for Opus 4.6, Sonnet 4.6, Opus 5.5 and Sonnet 5.5 mark Mumbai and Hyderabad as Global cross-Region only, and AWS says Global routing can reach any Region worldwide. Anthropic's table lists an apac profile for Opus 4.6, but AWS's card shows only US, EU and Australia, so call GetInferenceProfile from your source Region before you rely on either. On Foundry, both 4.6 models run on Anthropic's infrastructure outside Azure, so re-read your DPDP processing records during the move.

US. The Claude API offers US-only inference at 1.1x prices. Microsoft's Claude page lists Data Zone Standard (US) for Azure-hosted claude-opus-5-5 and claude-sonnet-5-5 but not for the 4.6 models, so the move can open a US-only option there, also at 1.1x.

EU. AWS lists eu profiles for both 4.6 models and for Opus 5.5 and Sonnet 5.5. Regional profiles carry a 10 percent premium over global endpoints, so budget for it.

Migration checklist with dates

Score each workflow before and after, including cost per successful task with thinking tokens. An AI gateway such as LiteLLM, Portkey, Kong or TrueFoundry makes cut-over and rollback a config change, and our guide to catching silent agent failures in CI/CD covers the gate.

How eCorpIT can help

eCorpIT is a Gurugram-based technology organisation, founded in 2021, assessed at CMMI Level 5 and MSME certified, with senior-led engineering teams working across AWS, Microsoft and Google platforms. Our AI evals and observability service builds golden sets from your production traces, compares Sonnet 5.5 and Opus 5.5 at several effort levels and gates the switch in CI. Talk to us at /contact-us/.

Last updated: 1 November 2026.

References

1. Model deprecations — Claude Platform Docs

1. Model retirement schedule, last updated 23 September 2026 — Microsoft Learn

1. Anthropic Claude Opus 4.6 model card — Amazon Bedrock User Guide

1. Anthropic Claude Sonnet 4.6 model card — Amazon Bedrock User Guide

1. Model lifecycle (Legacy) — Amazon Bedrock User Guide

1. Plans and pricing — Claude

1. Claude Sonnet 5.5 migration guide — Claude Platform Docs

1. Claude Opus 5.5 migration guide — Claude Platform Docs

1. Foundry Models lifecycle and support policy, last updated 24 July 2026 — Microsoft Learn

1. Partner model deprecations, last updated 30 October 2026 — Google Cloud

1. Models overview — Claude Platform Docs

1. List models, API reference — Claude Platform Docs

1. Pricing — Claude Platform Docs

1. Claude Opus 5.5 — Claude Platform Docs

1. Anthropic Claude Opus 5.5 model card — Amazon Bedrock User Guide

1. Anthropic Claude Sonnet 5.5 model card — Amazon Bedrock User Guide

1. Claude Platform release notes, entries of 21 and 30 October 2026 — Anthropic

1. Thinking — Claude Platform Docs

1. anthropic Python package, release history — PyPI

1. Claude on Amazon Bedrock (Opus 4.6 and earlier) — Claude Platform Docs

1. Migrating to API version 2026-10-21 — Claude Platform Docs

1. Claude models in Microsoft Foundry — Microsoft Learn