Originally published: 2026-10-01 14:08 UTC / 16:08 Berlin / 10:08 EDT Last verified: 2026-10-01 14:08 UTC No corrections at this time.
On September 30, 2026, Anthropic formally deprecated claude-sonnet-4-5-20250929 and set a hard retirement date of November 30, 2026. Requests to the model past that date will fail. The recommended replacement is claude-sonnet-5-5 (no date suffix). The deprecation is documented on the Anthropic platform release notes (Sep 30 entry), the model deprecations page, and the dedicated Sonnet 5.5 migration guide.
The Sep 30 announcement is exactly the kind of operational change the desk tracks: a 60-day forced migration window for a model that is widely used in production Claude API stacks. Today is day 1 of that window. The window is not negotiable — the deprecations page is explicit: "Requests to models past the retirement date will fail."
This article is documentation surfacing, not a firsthand test. All quoted lines are pulled directly from the official sources listed in the Sources section at the end. No API call was placed against claude-sonnet-4-5-20250929 or claude-sonnet-5-5 for this report; verification level is documentation comparison plus cross-reference to the existing mr.technology Sonnet 5.5 launch coverage.
1. The model status row. From the model deprecations page (verified 2026-10-01 14:08 UTC, verbatim): the table row for claude-sonnet-4-5-20250929 reads Deprecated | September 30, 2026 | November 30, 2026. The deprecation history section then states: "On September 30, 2026, Anthropic notified developers using Claude Sonnet 4.5 of its upcoming retirement on the Claude API." The recommended replacement is claude-sonnet-5-5.
2. The platform release note. The Sep 30, 2026 entry on the platform release notes overview (verbatim): "We announced the deprecation of the Claude Sonnet 4.5 model (claude-sonnet-4-5-20250929), with retirement on the Claude API scheduled for November 30, 2026. We recommend migrating to Claude Sonnet 5.5. Read more in Model deprecations." The announcement applies on Anthropic-operated platforms (Claude API, Claude Platform on AWS, Microsoft Foundry). Partner-operated platforms (Amazon Bedrock, Google Cloud) set their own retirement schedules.
3. The migration checklist. The Sonnet 5.5 migration guide has a dedicated "Claude Sonnet 4.5 or earlier" section with these changes (verbatim from the guide):
computer_20250124 to computer_20251124.output_config.effort explicitly.interleaved-thinking-2025-05-14, and replace fine-grained-tool-streaming-2025-05-14 with eager_input_streaming.output_format to output_config.format.The guide also notes that forced tool use (tool_choice: {type: "any" | "tool"}) returns a 400 error on Claude Sonnet 5.5. Migrating agents must use tool_choice: {type: "auto"} and mark the tool strict: true. On Amazon Bedrock, strict tool use is not available, so the prompt must explicitly say when to call the tool and the application must validate the input.
4. The pricing delta. Verified on the Anthropic pricing page (2026-10-01 14:08 UTC). Standard tier, per million tokens:
| Model | Input | 5-min cache write | 1-h cache write | Cache read | Output |
|---|---|---|---|---|---|
| Claude Sonnet 4.5 (deprecated) | $3.00 | $3.75 | $6.00 | $0.30 | $15.00 |
| Claude Sonnet 5.5 (replacement) | $2.00 | $2.50 | $4.00 | $0.20 | $10.00 |
The replacement is 33% cheaper on input, 33% cheaper on output, and 33% cheaper on cache reads. Batch tier drops from $1.50/$7.50 (4.5) to $1.00/$5.00 (5.5) for the same -33% delta. The 60-day migration is also a forced cost-out for every customer still calling the deprecated ID.
For developers maintaining integration code: the migration is not a model-ID swap. Three Sonnet-4.5-isms break on Sonnet 5.5 with a 400 error: assistant prefills, tool_choice: any and tool, and thinking: {"type": "disabled"} at non-between_tools levels. The migration guide has explicit before/after request bodies for each. Any agent that calls a tool with tool_choice: {type: "tool"} to force invocation will silently break on November 30. Plan the refactor against the checklist now.
For founders running cost-sensitive workloads: the migration is a 33% cost reduction across the rate card. A team spending $100k/month on Sonnet 4.5 standard tier input drops to $67k/month on Sonnet 5.5 with identical output behavior on the same task mix. The cost-out is automatic once the migration lands. It also closes the cache-read gap: at $0.20/MTok, Sonnet 5.5 is on the same cheap cache-read tier as Opus 5.
For platform owners managing Enterprise tier deployments: the deprecation policy states Anthropic provides at least 60 days' notice for publicly released models. The Sep 30 → Nov 30 window is 61 days, which meets the policy. The deprecations page has an export-from-Usage-page CSV workflow specifically designed to find residual claude-sonnet-4-5-20250929 traffic: Usage → Export → filter by model. That export should run in the first week of October.
For operators running Bedrock or Vertex AI: the dates differ. Anthropic explicitly notes that Amazon Bedrock and Google Cloud set their own retirement schedules. A model on Anthropic's Claude API retiring November 30 may still be served on Bedrock or Vertex into 2027. Confirm the partner-side date before scheduling the migration. The Sonnet 4.5 Bedrock ID is anthropic.claude-sonnet-4-5-20250929-v1:0; on Vertex it is claude-sonnet-4-5@20250929.
Primary source 1 — Anthropic platform release notes (Sep 30, 2026 entry) (verified verbatim at fetch 2026-10-01 14:08 UTC): <https://platform.claude.com/docs/en/release-notes/overview>
Primary source 2 — Model deprecations page (verified verbatim at fetch 2026-10-01 14:08 UTC): <https://platform.claude.com/docs/en/about-claude/model-deprecations>. Confirms the deprecation date (Sep 30), retirement date (Nov 30), recommended replacement (claude-sonnet-5-5), and the at-least-60-day notification policy.
Primary source 3 — Claude Sonnet 5.5 migration guide (verified verbatim at fetch 2026-10-01 14:08 UTC): <https://platform.claude.com/docs/en/models/sonnet-5-5/migration-guide>. The "Claude Sonnet 4.5 or earlier" section contains the seven migration checklist items quoted above.
Primary source 4 — Anthropic pricing page (verified verbatim at fetch 2026-10-01 14:08 UTC): <https://platform.claude.com/docs/en/about-claude/pricing>. The standard-tier and batch-tier rate rows for Sonnet 4.5 and Sonnet 5.5 are quoted in the table above.
Cross-check — mr.technology prior coverage: the Sep 29 03:33 UTC article claude-sonnet-5-5-launch-five-breaking-changes-account-bound-thinking-sep-2026 covers the model-side migration surface for Sonnet 5.5 but did not have the Sonnet 4.5 deprecation announcement because the deprecation was published one day after that article was filed. This article is the deprecation announcement, not a re-cap of the launch.
Verification level: documentation comparison against four primary Anthropic pages. No API calls placed. No first-party migration executed. The 33% pricing delta is arithmetic from the pricing page rate rows, not a measured workload.
Cost. The migration is a 33% cost reduction across standard and batch tiers on input, output, and cache reads. At aggregate scale, this is a meaningful recurring cost-out for any team that has not already migrated. Cache reads drop from $0.30 to $0.20 per MTok — relevant for prompt-caching-heavy workloads. The pricing table above is the source of truth; the cost math is straightforward.
Risk. Three classes of break risk for migration:
1. Silent 400 errors for tool_choice: any/tool on the new model. A code path that worked yesterday on Sonnet 4.5 will return a 400 on Sonnet 5.5. The migration guide's "Forced tool use is not supported" section is explicit about the 400 and the auto+strict replacement. 2. Assistant prefill removal. A code path that relied on prefills to seed the model's behavior will degrade on Sonnet 5.5. Migration is to remove the prefill and reword the prompt to elicit the same behavior. 3. Thinking parameter shape change. thinking: {type: "disabled"} returns a 400 on Sonnet 5.5; the replacement is thinking: {type: "between_tools"} at low, medium, or high effort, or thinking: {type: "adaptive"} at xhigh or max. Any code path that hard-coded "type": "disabled" to suppress thinking will need to switch.
Risk on partner-operated platforms. Amazon Bedrock and Google Cloud keep their own retirement schedules. The Anthropic-operated Claude API retirement on November 30 does not guarantee the same date on Bedrock or Vertex. Do not assume a single calendar date across platforms.
Limitations of this report. This is documentation surfacing. No live migration was executed against a production Sonnet 4.5 workload. The 33% cost reduction is derived arithmetically from the published rate rows; a measured workload comparison (Sonnet 4.5 vs Sonnet 5.5 on the same prompts at the same effort) is the next step if you depend on the absolute numbers. The migration guide's before/after request examples are quoted from the documentation, not validated against a live API.
This is the standard deprecation playbook — Anthropic has been running it since 2024 — but it lands on a Sonnet 4.5 → Sonnet 5.5 step that is genuinely a 33% cost reduction with no capability regression at the headline level. The 60-day window is the policy minimum, which means there is no slack. Teams that have not started the migration by mid-October will be in a forced-migration sprint at the start of November.
The deprecation is also a forcing function for the tool_choice: any/tool pattern. Many agent frameworks use forced tool invocation to bypass the model's refusal to call. That pattern is a 400 on Sonnet 5.5. The migration to auto + strict: true is the right direction (it forces schema validation), but it is a non-trivial refactor for any code path that depended on the old shape. Plan the refactor against the checklist in the migration guide, not as a model-ID swap.
For teams that already use Sonnet 5.5 (because of the Sep 28 launch), this announcement is a non-event. For teams still calling claude-sonnet-4-5-20250929, it is a 60-day clock that ends in forced failure on December 1, 2026.
Today (Oct 1):
1. Export a CSV from the Claude Console Usage page filtered by model=claude-sonnet-4-5-20250929. Identify every internal code path still calling the deprecated model ID. Note the volume of traffic and which teams own each path. 2. Read the Sonnet 5.5 migration guide "Claude Sonnet 4.5 or earlier" section end-to-end. The seven checklist items above are the minimum code-side surface. 3. Pin a single owner per code path to land the migration. The forced-tool-use and assistant-prefill changes are non-trivial; they are not a model-ID swap.
This week (Oct 1-7):
4. Stand up a staging environment running claude-sonnet-5-5 behind the same code path that currently calls Sonnet 4.5. Run the same prompt suite against both models. Confirm the output is acceptable and the cost delta matches the rate-card arithmetic. 5. For Bedrock or Vertex deployments, contact the platform owner for the partner-side retirement schedule. The Anthropic November 30 date is not guaranteed on Bedrock or Vertex. 6. If you have agent code that uses tool_choice: {type: "tool"} to force invocation, refactor to auto + strict: true. Validate the input shape against the tool schema before passing it to the tool.
By Oct 31:
7. Cut all production traffic from claude-sonnet-4-5-20250929 to claude-sonnet-5-5 (or the partner-side equivalent). Allow one week of overlap on a controlled canary.
By Nov 29 (one day before retirement):
8. Verify zero production traffic on the deprecated model ID. Confirm the cutover audit matches the Oct 1 export.
Skip if not in scope: if your code only calls Sonnet 5.5, Opus 5.5, Fable 5.1, Mythos 5.1, or any other Active model on the deprecations page, this announcement does not affect you. The deprecation is specific to claude-sonnet-4-5-20250929.
claude-sonnet-4-5-20250929 or claude-sonnet-5-5 for this report. The 33% pricing delta is arithmetic from the published rate rows. The migration checklist items are quoted verbatim from the Sonnet 5.5 migration guide. The deprecation row in the deprecations table is quoted verbatim.