Guide from the official docs
Claude Sonnet 4.5 retires Nov 30: will your OpenClaw, Hermes or Claude Code setup break?
Claude Sonnet 4.5 stops working on the Claude API and Microsoft Foundry on Nov 30, 2026. Where the model name hides in OpenClaw, Hermes Agent and Claude Code, and what to change.
This guide follows the official docs and announcements, linked inline. AideDex did not test these products and does not score them. How we check facts.
Last checked Nov 13, 2026.
If your personal agent calls Claude through Anthropic’s own API, Microsoft Foundry, Claude Platform on AWS or Claude on Google Cloud (Anthropic managed), and its settings name claude-sonnet-4-5 or claude-sonnet-4-5-20250929, it stops working on Nov 30, 2026. Anthropic’s deprecation page says requests to a retired model fail and names claude-sonnet-5-5 as the replacement (Anthropic, Microsoft Foundry schedule). Amazon Bedrock runs its own, later clock. Google’s own page for Claude Sonnet 4.5 on Google Cloud says the same Nov 30 shutdown applies there, to existing customers only (Google Cloud). Nothing in the docs we read says the model moves to the new one for you.
The write-ups we found for this retirement are aimed at developers with an application. This one is for people who run an agent such as OpenClaw or Hermes Agent, or use Claude Code, and want to know where the old model name is stored on their own machine.
The dates by platform
| Where Claude is called | Sonnet 4.5 | Who sets the date |
|---|---|---|
| Claude API | Deprecated Sep 30, 2026; retires Nov 30, 2026 | Anthropic (deprecations page, endoflife.ai) |
| Claude Platform on AWS, Claude on Google Cloud (Anthropic managed) | Same date, Nov 30, 2026 | Anthropic says its page covers these (deprecations page, Since.dev) |
| Microsoft Foundry | Nov 30, 2026, replacement claude-sonnet-5-5 | Listed on Microsoft’s own schedule, version 1 (Microsoft, Anthropic) |
| Claude on Google Cloud (Google managed) | Deprecated Sep 29, 2026; shut down Nov 30, 2026; existing customers only | Google (Google Cloud page, updated Nov 12, 2026) |
| Amazon Bedrock | Legacy since Oct 8, 2026; extended access from Jan 8, 2027; end of life Apr 8, 2027 (commercial Regions; AWS says dates can differ by Region, and GovCloud runs Oct 27, 2026 / Jan 27, 2027 / Apr 27, 2027) | AWS (AWS lifecycle page, Since.dev) |
Google publishes its own date for the Google-managed offering, and it matches Anthropic’s: Nov 30, 2026 (Google Cloud). Anthropic’s page says Google sets this schedule itself, so if your agent calls Claude there, check Google’s page for your project before the date.
On Bedrock, AWS adds three conditions that matter for a hobby agent. New customers cannot use a Legacy model, existing customers may lose access after 15 days of inactivity, and during the extended-access period you should expect higher pricing set by the model provider (AWS, Since.dev). An agent you only switch on once a month could lose access before April.
Two names for the same model
Anthropic’s model ID page says models before the 4.6 generation carry a snapshot date, and that on the Claude API they also have a short alias: claude-sonnet-4-5 points to the most recent dated snapshot for that minor version (Anthropic, Since.dev). So a config that says claude-sonnet-4-5 is still Sonnet 4.5. Search your files for both spellings, and for sonnet-4.5 with a dot, since people type it that way.
Where the name can hide in your agent
OpenClaw
OpenClaw’s agents.defaults.model takes either a plain provider/model string or an object with a primary and a list of fallbacks (OpenClaw config docs). To see what your install resolved, the CLI reference lists openclaw models status and openclaw models list (OpenClaw CLI reference).
Four details in the docs decide whether a retired model hurts you:
- A fallback rescues you and a missing one does not. The failover page says that when an error explicitly identifies an unknown, missing or retired model, OpenClaw moves to the configured fallback, and without one it surfaces the error (OpenClaw failover). It also treats an HTTP 404
model_not_foundas a reason to fall back, though a 404 whose body points to some other condition does not count (OpenClaw failover). - A model you picked by hand in a chat is strict. Choosing a model other than the default with
/modelcreates a user selection. If that model becomes unreachable, the docs say the run fails visibly instead of answering from another model, even when fallbacks exist (OpenClaw Models CLI, OpenClaw failover). The Models page also says changing the global primary does not rewrite existing session pins;/model defaultclears one (OpenClaw Models CLI). - Scheduled jobs carry their own model. A cron job’s
--modelis a per-job primary that uses the configured fallbacks unless the job sets its own list (OpenClaw Models CLI, OpenClaw failover). An old nightly job is the kind of place a forgotten model name sits. - An agent with its own model needs its own fallback. In the failover docs, a native agent’s
agents.entries.*.modelis strict unless that agent’s model object lists its ownfallbacks, so the list underagents.defaults.modeldoes not cover it (OpenClaw failover).
Hermes Agent
Hermes keeps the main model in the model: block of ~/.hermes/config.yaml (hermes config edit opens it) (Hermes configuration); the provider docs show provider: "anthropic" with default: "claude-sonnet-4-6" as their example, and hermes model is the setup wizard while /model inside a chat switches between providers you already configured (Hermes providers). The example uses Sonnet 4.6, not 4.5, so it says nothing about what your file holds; open it and look.
Two more places can name a model. The configuration docs describe auxiliary: blocks for side tasks such as vision and compression, which follow your main model by default but can each be pointed at a different one, and a top-level fallback_providers: list for the main chain (Hermes configuration, Hermes providers). If an auxiliary task was set to a Sonnet 4.5 ID by hand, it will not follow your main model when you change that.
Claude Code
The sonnet alias in Claude Code does not mean the same model everywhere. Claude Code’s model page lists what each alias resolves to by provider: on the Anthropic API sonnet is Sonnet 5.5, but on Amazon Bedrock, Google Cloud’s Agent Platform and Microsoft Foundry it is Sonnet 4.5 (Claude Code model config, Claude Code on Bedrock). The Bedrock page confirms the Bedrock case, and both the Bedrock and Foundry pages tell you to pin versions yourself because the built-in default can lag the newest release (Claude Code on Bedrock, Claude Code on Foundry).
Put those two sets of dates together and the Foundry case stands out. By those docs, someone who runs Claude Code on Foundry with no pin is on Sonnet 4.5, both through the sonnet alias and through the default model Claude Code uses when you set none, and Microsoft’s schedule ends Sonnet 4.5 on Nov 30 (Claude Code model config, Microsoft). The Foundry page adds that Claude Code has no startup model check there, so requests fail when the default is unavailable (Claude Code on Foundry). We did not test what happens on Nov 30, so treat this as a reading of three documents and check your own deployment. The fix the docs describe is to set ANTHROPIC_DEFAULT_SONNET_MODEL to a model you choose (Claude Code model config, Claude Code on Foundry).
Google Cloud’s Agent Platform is different in one way. Claude Code’s docs say the sonnet alias resolves to Sonnet 4.5 there unless you set ANTHROPIC_DEFAULT_SONNET_MODEL, and Google’s page puts Sonnet 4.5’s shutdown at Nov 30; but the same docs say Claude Code checks models at startup and falls back to an earlier or lower-tier model when the default is unavailable, so a silent downgrade is more likely than a hard failure (Claude Code on Google Cloud’s Agent Platform, Google Cloud). The Claude Code page does not say which Claude offering it calls there, so confirm that in your own project. We did not test this.
On Bedrock the same default keeps working past Nov 30, because AWS ends the model on Apr 8, 2027, but it is already on AWS’s Legacy clock.
What changes when you move to Sonnet 5.5
Changing the name is not the whole job if your tool sends raw API settings. Anthropic’s migration guide lists several changes from Sonnet 4.5: thinking runs by default, prefilled assistant turns and temperature, top_p and top_k return a 400 error, and the same text produces about 30 percent more tokens (Anthropic migration guide, GrtLabs). For an agent you did not write, the agent’s own developers deal with the request settings; what you may notice is a different bill and different speed. Our hosting and cost guide lists Sonnet 5.5 prices, and the extra tokens would count on top of them.
The model you move to matters, too. Anthropic lists Sonnet 4.6 as active with a retirement date of not sooner than Feb 17, 2027, and Microsoft lists Sonnet 4.6 on Foundry for Feb 10, 2027 (Anthropic, Microsoft). Moving there is a smaller step but buys a few months, so it is a choice you are making on purpose.
Other Claude 4.5 dates on Foundry
The same Microsoft table lists two more Claude 4.5 models: Opus 4.5 with a retirement date of Nov 24, 2026 and Haiku 4.5 with Nov 15, 2026, both with no replacement named (Microsoft). Anthropic’s own page still shows both as active, with “not sooner than” dates of Nov 24, 2026 for Opus 4.5 and Oct 15, 2026 for Haiku 4.5 and no deprecation notice (Anthropic). We cannot tell you why the two pages differ. If you use those models on Foundry, read Microsoft’s table for your own deployment.
What to do before Nov 30
- Search your agent’s config files, environment variables, cron jobs and any saved sessions for
sonnet-4-5,sonnet-4.5andclaude-sonnet-4-5-20250929. - If you use OpenClaw (our setup guide covers the install), run
openclaw models statusand look at the primary and fallbacks, check any agent that sets its ownmodel, and add a fallback that is not Sonnet 4.5.openclaw config get agents.defaults.model --jsonprints the default block (OpenClaw CLI reference). - If you use Hermes Agent (see our setup guide), open
~/.hermes/config.yamland check themodel:block, everyauxiliary:entry andfallback_providers:. - If you use Claude Code through Foundry, Bedrock or Google Cloud, pin the Sonnet model yourself instead of relying on the alias. On Foundry and Google Cloud that matters before Nov 30. For teams, the managed settings guide covers where an admin sets Claude Code policy.
- Pick the replacement deliberately: Sonnet 5.5 is the one Anthropic and Microsoft name, and a retirement date can still move for models marked “not sooner than”.
- After you change the model, if the agent errors on startup, our setup errors guide lists the common OpenClaw and Hermes failures, and Is my AI agent down? helps separate an outage from your own config.
What we checked and what we did not
On Nov 13, 2026 we read the pages linked above: Anthropic’s deprecation, model ID and migration pages, Microsoft’s and AWS’s lifecycle tables, Google Cloud’s Claude Sonnet 4.5 page, the Claude Code docs for model configuration, Bedrock, Foundry and Google Cloud’s Agent Platform, and the OpenClaw and Hermes Agent docs. We did not run OpenClaw, Hermes Agent or Claude Code against a retired model, so what happens on Nov 30 inside each tool comes from the docs only. We also did not find any statement about whether the Claude apps (web, phone, desktop) change their model list on Nov 30, because the dates above cover API platforms; if you only use the apps, watch the model picker rather than this table.
FAQ
Will my agent switch to Sonnet 5.5 by itself?
Only if you set it up that way. Anthropic’s page says requests to models past the retirement date will fail, and AWS says migration on Bedrock does not happen automatically (Anthropic, AWS). OpenClaw can move to a fallback you have set, but only if you set one (OpenClaw failover).
I use an agent that never shows me a model name. Am I affected?
We cannot say. The retirement notice names API model IDs, and a tool that hides its model gives you nothing to search for. If a tool lets you pick “Claude Sonnet 4.5” from a menu, or you do not know which model it calls, ask its vendor in writing what happens on Nov 30 (GrtLabs).
Is there a deadline to test the new model?
Anthropic recommends testing a replacement well before the retirement date, and its page says customers with active deployments are notified at least 60 days ahead for publicly released models (Anthropic, endoflife.ai). Sonnet 4.5 was deprecated on Sep 30 and retires on Nov 30, so after Nov 30 you cannot go back to it on the Claude API.
Sources
19 pages- Claude Platform Docs: Model deprecations (read Nov 13, 2026) platform.claude.com
- Claude Platform Docs: Model IDs and versioning (read Nov 13, 2026) platform.claude.com
- Claude Platform Docs: Claude Sonnet 5.5 migration guide (read Nov 13, 2026) platform.claude.com
- Microsoft Learn: Model retirement schedule for Microsoft Foundry (read Nov 13, 2026) learn.microsoft.com
- AWS: Amazon Bedrock model lifecycle, Legacy (read Nov 13, 2026) docs.aws.amazon.com
- Claude Code Docs: Model configuration (read Nov 13, 2026) code.claude.com
- Claude Code Docs: Amazon Bedrock (read Nov 13, 2026) code.claude.com
- Claude Code Docs: Microsoft Foundry (read Nov 13, 2026) code.claude.com
- Claude Code Docs: Google Cloud's Agent Platform (read Nov 13, 2026) code.claude.com
- Google Cloud Docs: Claude Sonnet 4.5 on Google Cloud (read Nov 13, 2026) cloud.google.com
- OpenClaw Docs: Models CLI (read Nov 13, 2026) docs.openclaw.ai
- OpenClaw Docs: Model failover (read Nov 13, 2026) docs.openclaw.ai
- OpenClaw Docs: Configuration, agent models (read Nov 13, 2026) docs.openclaw.ai
- OpenClaw Docs: openclaw models CLI reference (read Nov 13, 2026) docs.openclaw.ai
- Hermes Agent Docs: LLM and Model Providers (read Nov 13, 2026) hermes-agent.nousresearch.com
- Hermes Agent Docs: Configuration (read Nov 13, 2026) hermes-agent.nousresearch.com
- endoflife.ai: Claude Sonnet 4.5 retires November 30, 2026 (updated Nov 13, 2026) endoflife.ai
- Since.dev: Claude Sonnet 4.5 retires on November 30, 2026 (updated Nov 4, 2026) since.dev
- GrtLabs: Claude Sonnet 4.5 retirement checklist (Nov 4, 2026) grtlabs.com