Claude Sonnet 4.5 stops serving API requests after November 30 with no grace period

Anthropic sets a hard cutoff

Anthropic moved Claude Sonnet 4.5 from Active to Deprecated status on September 30, 2026, according to a post on dev.to that walks through the notice. Any request to the exact model id claude-sonnet-4-5-20250929 will be rejected after November 30, 2026, with nothing in the way of a grace period once that date passes. The gap between the deprecation notice and the shutoff is 61 days, one day beyond the minimum 60 days of notice Anthropic has publicly committed to giving for its released models, the post points out.

The named replacement is claude-sonnet-5-5, which the dev.to article says carries identical pricing to the outgoing model and the same 1 million token context window, so the migration is not driven by cost or capacity changes.

This is not an isolated event. Sonnet 4.5 is the sixth Anthropic model to receive a deprecation notice since October 2025, following Claude Opus 4.1 in June and Claude Sonnet 4 and Claude Opus 4 together in April, according to the post. The lead time has varied: Claude Sonnet 3.7 was flagged on October 28, 2025 and retired on February 19, 2026, nearly four months later, while Claude Sonnet 4 got roughly two months between notice and retirement.

One deadline does not cover every platform

A detail worth flagging: Anthropic's deprecation page applies only to Anthropic-operated platforms, namely the Claude API, Claude Platform on AWS and Microsoft Foundry. Amazon Bedrock and Google Cloud maintain their own retirement schedules for the same underlying models, so teams calling Sonnet 4.5 through Bedrock cannot assume the November 30 date binds them and should consult that provider's model table instead, the dev.to author cautions.

Dated ids are the real exposure

The practical risk, as the post frames it, is less about the model itself and more about the literal string claude-sonnet-4-5-20250929 sitting in a config file, an environment variable or a script nobody has opened since it was written. Anthropic issues two kinds of model names: dated snapshot ids that pin one exact version forever, and rolling aliases that follow whatever version is current. A rolling name like claude-sonnet-5-5 keeps working through version changes, while a dated id is frozen and can be deprecated out from under a running system. The author notes that environment variables such as ANTHROPIC_DEFAULT_OPUS_MODEL, when set to a snapshot id and forgotten, fail the same way: silently fine until the day the model behind them disappears.

How to find affected calls before the deadline

The fastest first step, according to the post, is the Export button on the Usage page in Claude Console, which produces a CSV broken down by API key and model. That shows exactly which keys are still hitting the retiring id before anyone has to grep through old code.

The swap itself is small, but the author warns against a blind find-and-replace. In a separate write-up covering five breaking changes when Sonnet 5.5 shipped on September 28, they found at least one change that alters output shape rather than raising an error, meaning a quick smoke test can pass while production behavior still differs. Beyond the id itself, they suggest checking retry and fallback logic that names models directly, cached prompts or few-shot examples tuned to Sonnet 4.5's phrasing habits, and any README or setup script that hardcodes the dated id, since those copies are the easiest to miss.

Retired does not mean deleted

Separately from the notice, Anthropic has a published commitment on deprecation and preservation, which the post summarizes: retired model weights are kept long-term rather than deleted, with the stated goal of eventually making past models accessible again. The same document acknowledges the tradeoffs, including lost access for researchers running comparisons, forced migration for anyone who depends on a model's exact behavior, and open safety and model-welfare questions around retiring models.

Why it matters

This is a fixed, dated shutoff with no grace period, and the failure mode lands on the systems least likely to be monitored: one-off scripts, pinned environment variables and setup docs written months ago. Checking costs almost nothing today, while waiting until December 1 turns a config string into an outage. Anthropic's pattern of a named replacement, a published lifecycle status and a minimum 60-day notice makes the migration predictable, but only for teams that actually read the notice and audit their usage before the deadline arrives.