Claude Sonnet 4.5 Retires on 30 November 2026: What to Do If Your App or AI Pilot Runs on It
Anthropic retires claude-sonnet-4-5-20250929 on 30 November 2026 and has removed temperature, top_p and top_k in API version 2026-10-21. The four-step migration for a 10 to 100 person company, what it costs with and without an engineer, and why a retirement is the moment to add the production layer.
Oct 30, 2026
If any part of your software calls the model `claude-sonnet-4-5-20250929`, it stops working on 30 November 2026. Anthropic's model deprecations page, read on 27 October 2026, lists Claude Sonnet 4.5 as deprecated since 30 September 2026 with a retirement date of 30 November 2026 and `claude-sonnet-5-5` as the recommended replacement. Requests to a retired model fail. The same page records a second change that catches older code: the `temperature`, `top_p` and `top_k` parameters were removed in API version 2026-10-21 and now return a 400 error, so a migration that only swaps the model name can still break. For a 10 to 100 person company this is a four-step job: find every call, switch the model, re-run your evaluation set, and remove the dead parameters. If you have an engineer it is a week of their time; if you do not, it sits inside KUMO's production layer work from $15,000, or inside an ongoing engineering retainer at $5,000 to $10,000 per month.
Key facts (read from Anthropic's model deprecations page on 27 October 2026)
- `claude-sonnet-4-5-20250929`: deprecated 30 September 2026, retires 30 November 2026, recommended replacement `claude-sonnet-5-5`
- Anthropic gives at least 60 days' notice before a public model retires and emails customers with active deployments; the notice for Sonnet 4.5 went out on 30 September
- API version 2026-10-21 removed `temperature`, `top_p` and `top_k`; requests that set them on that version return a 400 error
- Other active models carry "not sooner than" dates: Claude Haiku 4.5 not sooner than 15 October 2026, Claude Opus 4.5 not sooner than 24 November 2026, so a migration plan should name the next date too
- Partner platforms (Amazon Bedrock, Google Cloud's Google-managed models) set their own retirement schedules; check the one you use
Who this affects
Anyone whose product, internal tool, support agent or document pipeline was built in the last year with Claude Sonnet 4.5 as the model, which includes a large share of AI pilots started in late 2025 and early 2026 and many apps built with Lovable, Bolt, Replit or n8n that pasted a model name into a configuration field and never looked again. The people who built the pilot may have moved on; the model name is in a file nobody has opened since.
The four steps
- Find every call. Anthropic's Console has a Usage page with an Export that lists usage by API key and model; filter for `claude-sonnet-4-5`. Then search your repositories, n8n or Make workflows, Lovable or Bubble settings and environment variables for the same string. The export finds what is running; the search finds what is dormant and will fail on first use.
- Switch the model and read the migration guide. The recommended replacement is `claude-sonnet-5-5`. Behaviour differs between model generations: output length, formatting, tool-call shape and refusals move. Anthropic publishes a migration guide for this; follow it rather than assuming a name swap is enough.
- Re-run your evaluation set. If you have one, run it against the new model before any customer does and compare the pass rate. If you do not have one, this is the moment it pays for itself: twenty to fifty real inputs with the answers you expect, checked by a person once, then automatically on every change. The AI pilot to production page explains why this set is the first thing a production layer adds.
- Remove the dead parameters. If your code sets `temperature`, `top_p` or `top_k` and you move to API version 2026-10-21, those calls return 400. Remove the parameters and guide behaviour through the prompt instead, as Anthropic's page advises. Then fix the next retirement date in your calendar.
What it costs
| Situation | What is needed | Cost |
|---|---|---|
| You have an engineer and an evaluation set | The four steps above | About a week of their time; no external cost |
| You have an engineer, no evaluation set | The four steps plus building the set from real cases | One to two weeks of their time |
| No engineer, one pilot or app on the model | KUMO runs the four steps and leaves you an evaluation set, a cost cap and a rollback | Inside the production-layer band, from $15,000 (4 to 6 weeks) |
| No engineer, several systems on the model | Migration across systems with monitoring and a fallback model | Inside the $20,000 to $50,000 band, or an ongoing engineering retainer at $5,000 to $10,000 per month |
KUMO bands are KUMO's published prices for the build only; model usage and hosting sit on top.
Why this is the right moment to add the production layer
A retirement forces the work that a pilot skipped: knowing where the model is called, having a set of cases that proves it still works, and being able to roll back. Doing the migration without those three means finding out from customers on 1 December. Doing it with them means the next retirement (every active model on Anthropic's page carries a "not sooner than" date) is an afternoon, not an incident. The production readiness check covers the same discipline for the app around the model.
What a KUMO engagement includes, and the proof
Included: the usage audit across Console and code, the migration with the published guide, an evaluation set from your real cases, a fallback to a second model where it matters, a cost cap, a deploy pipeline with rollback, and handover of everything to accounts you control. The proof is direct for the production discipline: CampaignHQ is KUMO's own product in production since 2022; adjacent: KUMO's founders built Volopay's first production version as its founding engineers. KUMO is a software engineering studio that builds custom AI agents, workflow apps and portals for 10 to 100 person companies, and a Registered Partner in the OpenAI Partner Network; see Built with Anthropic Claude for the Claude work and AI integration for how model migrations are scoped.
How to get a real number this week
Write down which systems call the model, whether you have an evaluation set, and who owns the code. With that, get a scoped estimate and KUMO returns a band and a migration plan in two working days, or book a 30-minute call with an engineer. If you only want to know whether you are affected, step 1 above takes ten minutes.
FAQ
When does Claude Sonnet 4.5 stop working? 30 November 2026, per Anthropic's model deprecations page read on 27 October 2026. Requests to `claude-sonnet-4-5-20250929` fail after that date on Anthropic-operated platforms.
What replaces Claude Sonnet 4.5? Anthropic recommends `claude-sonnet-5-5`. Test with your own cases before switching; behaviour differs between model generations.
Will my app break if I just change the model name? It may. API version 2026-10-21 removed `temperature`, `top_p` and `top_k`; calls that set them return a 400 error. Remove the parameters and re-run your evaluation set.
How do I find out if we use it? Export usage from the Anthropic Console Usage page and filter by model; then search your code, workflows and builder settings for the model name.
How much does a model migration cost if we have no engineer? From $15,000 inside KUMO's production-layer band for one pilot or app, including an evaluation set and a rollback; $20,000 to $50,000 across several systems; or an ongoing retainer at $5,000 to $10,000 per month.
Does this apply on Amazon Bedrock or Google Cloud? Partner-operated platforms set their own schedules; Anthropic's page says to check their model tables. The Anthropic API, Claude Platform on AWS, Anthropic-managed models on Google Cloud and Microsoft Foundry follow the dates above.
Last verified 27 October 2026 against Anthropic's model deprecations page (platform.claude.com). Dates are Anthropic's; check the page before you act, since Anthropic updates it. KUMO bands are KUMO's published prices.
