Claude Sonnet 4.5 has 61 days to live, and the replacement Anthropic names is 13% cheaper — not 33%
TL;DR: On 30 September 2026 Anthropic deprecated claude-sonnet-4-5-20250929, retiring 30 November 2026 — 61 days of notice against a published promise of “at least 60 days,” and the fifth deprecation in a row to land at 60, 61 or 62 days. The five before averaged 125 days. Sonnet 4.5 shipped 29 September 2025, so it dies at 14 months old. The named replacement, Claude Sonnet 5.5, lists at $2/$10 against Sonnet 4.5’s $3/$15 — a 33% cut that shrinks to about 13% once the newer tokenizer’s ~30% more tokens for the same text is counted. Meanwhile Haiku 4.5 sits in the current lineup with a retirement floor of 15 October 2026, which third-party trackers are misreading as a deadline. It is not one.
The notice period has converged on its own minimum
Anthropic’s deprecation documentation contains a promise worth quoting exactly: the company “notifies customers with active deployments for models with upcoming retirements, providing at least 60 days’ notice before model retirement for publicly released models.”
On 30 September 2026, Anthropic notified developers that claude-sonnet-4-5-20250929 retires on 30 November 2026. Count the days: 61. Taken alone that is unremarkable — a commitment honoured with a day to spare. The interesting number appears when the deprecation history is read as a series rather than a list of events.
| Announced | Model(s) | Retirement | Notice |
|---|---|---|---|
| 4 Sep 2024 | Claude 1.x, Instant 1.x | 6 Nov 2024 | 63 days |
| 21 Jan 2025 | Claude 2, 2.1, Sonnet 3 | 21 Jul 2025 | 181 days |
| 30 Jun 2025 | Opus 3 | 5 Jan 2026 | 189 days |
| 13 Aug 2025 | Sonnet 3.5 (both snapshots) | 28 Oct 2025 | 76 days |
| 28 Oct 2025 | Sonnet 3.7 | 19 Feb 2026 | 114 days |
| 19 Dec 2025 | Haiku 3.5 | 19 Feb 2026 | 62 days |
| 19 Feb 2026 | Haiku 3 | 20 Apr 2026 | 60 days |
| 14 Apr 2026 | Sonnet 4, Opus 4 | 15 Jun 2026 | 62 days |
| 5 Jun 2026 | Opus 4.1 | 5 Aug 2026 | 61 days |
| 30 Sep 2026 | Sonnet 4.5 | 30 Nov 2026 | 61 days |
The first five announcements average 124.6 days. The last five average 61.2. The policy did not change — the behaviour did. What was once a commitment with generous headroom is now a commitment met precisely, five times running, with a spread of two days.
Model lifespans tell the same story from the other end. Opus 3 served about 22 months before retirement and Haiku 3 about 25. The 4-series models are going dark at 12 to 14 — Sonnet 3.7 at exactly 12 months, Opus 4.1 at exactly 12, Sonnet 4 and Opus 4 at 13, and now Sonnet 4.5 at 14, deprecated one day after its first birthday.
For a buyer that converts a vague sense that “models get replaced” into two planning constants: roughly 12 to 14 months of service, and roughly 60 days of warning. Neither is in a contract. Both are now well-evidenced by the vendor’s own published record.
The replacement is named, and it is not the gentlest one
The deprecation table gives one recommended replacement for Sonnet 4.5: claude-sonnet-5-5. Three active Sonnets could take the traffic, and they differ in ways the single-cell recommendation flattens.
| Target | List price | Tokenizer | Lifecycle label | Retirement floor |
|---|---|---|---|---|
| Sonnet 4.6 | $3 / $15 | previous | Legacy | not sooner than 17 Feb 2027 |
| Sonnet 5 | $2 / $10 | current | Legacy | not sooner than 30 Jun 2027 |
| Sonnet 5.5 | $2 / $10 | current | Current lineup | not sooner than 28 Sep 2027 |
Sonnet 4.6 is the near-drop-in: identical rate card, the same older tokenizer, and — crucially — it still accepts the request shapes Sonnet 4.5 code uses. It is also already labelled legacy, with a floor only two and a half months past the retirement you are migrating away from. Choosing it means doing this exercise again inside a quarter.
Sonnet 5.5 is the right destination for most teams: the only one of the three in the current lineup, the longest floor, and a June 2026 knowledge cutoff seventeen months fresher than the model it replaces. The point is not that Anthropic recommended wrongly. It is that the recommendation is also the migration with the longest checklist — five documented breaking changes plus one that raises no error — presented as a single cell in a two-column table.
Two things break before anything model-specific does
The Sonnet 5.5 migration notes document the model-level changes. Two platform-level changes bite first, and both are easy to miss because they predate this deprecation entirely.
Sampling parameters are gone. temperature, top_p and top_k are deprecated from Claude Opus 4.7 onward and return a 400 error when set to a non-default value. The Python SDK removed them from its request types in v1.0, so passing them raises a TypeError before a request is even made. Virtually every integration written against Sonnet 4.5 in 2025 sets temperature. On Sonnet 4.6 those parameters still work; on Sonnet 5.5 they do not — which is both the strongest practical argument for the 4.6 bridge and the strongest argument against treating the swap as a config change.
Extended thinking is not accepted. The thinking.type: "enabled" plus budget_tokens mode Sonnet 4.5 used is deprecated on Opus 4.6 and Sonnet 4.6 and, per the models overview, “not accepted on later models.” Sonnet 5.5 uses adaptive thinking steered by an effort parameter instead — a different cost profile, and a default of high that Anthropic’s own agentic guidance steers away from.
The 33% price cut is about 13%
Sonnet 4.5 bills at $3 per million input tokens and $15 per million output. Sonnet 5.5 bills at $2 and $10 — a 33% reduction, and on a spreadsheet of rate cards it is one.
The pricing page carries a paragraph that changes the answer. Claude 4.7 and later models use a newer tokenizer that “produces approximately 30% more tokens for the same text,” and Sonnet 4.6 and earlier use the previous one. Sonnet 4.5 is on the old tokenizer; Sonnet 5.5 is on the new one. The unit of billing is not the same unit on both sides of the migration.
Hold the text constant and re-run it. A corpus that metered as one million tokens on Sonnet 4.5 meters as roughly 1.3 million on Sonnet 5.5:
| Line | Sonnet 4.5 | Sonnet 5.5 (tokenizer-adjusted) | Change |
|---|---|---|---|
| Input, per million old tokens | $3.00 | 1.3 × $2.00 = $2.60 | −13% |
| Output, per million old tokens | $15.00 | 1.3 × $10.00 = $13.00 | −13% |
| Cache read, per million old tokens | $0.30 | 1.3 × $0.20 = $0.26 | −13% |
The saving is real and consistent across all three line items. It is also less than half the headline. Anthropic is upfront that the exact increase “depends on the content and workload shape,” so 13% is a direction of travel rather than a quotable figure — which is precisely why the measurement has to happen on your own corpus before the number reaches a budget. What the arithmetic rules out is reading the rate card alone and booking a third off.
There is a real offset in the other direction. Sonnet 5.5 includes the full 1M-token context window at standard pricing, drops the minimum cacheable prompt to 512 tokens, and cuts tool-use system prompt overhead from 496 tokens to 286. A faster model that needs fewer tool calls can finish the same task for fewer tokens even at a 30% tokenizer penalty. That is a genuine saving — it is just a claim about your workload, not a property of the price list. Cache economics have become where these refreshes actually move money, and cache behaviour is workload-shaped by definition.
The floor that is being read as a deadline
One more row in the status table deserves attention, because the ecosystem is actively misreading it.
claude-haiku-4-5-20251001 is listed as Active, with Deprecated marked N/A and a tentative retirement date of “Not sooner than October 15, 2026.” That is 14 days from today. Haiku 4.5 is not a legacy model; it sits in the four-model current lineup as “the fastest model with near-frontier intelligence.”
Third-party lifecycle trackers and auto-generated repository issues have converted that floor into an end-of-life date, and teams are filing migration tickets against it. The reading is wrong twice over. No deprecation notice has been issued, and that notice is the event that starts a retirement clock. And the 60-day commitment makes a 15 October retirement arithmetically impossible — a notice published today would put the earliest retirement around 1 December.
The genuine signal is quieter. Haiku 4.5 is the only model in the current lineup whose floor falls inside the next twelve months; every other current model is committed into late 2027. Its reliable knowledge cutoff is February 2025, now twenty months stale. It is the obvious next candidate for a notice — a reason to have a Haiku migration plan written, not a reason to execute it this fortnight. The distinction between “deprecated” and “has a floor date” is the whole difference between an urgent project and a documented one.
The pattern across vendors
Anthropic is not an outlier, which is what makes the compression a market fact rather than a complaint: OpenAI is shutting down its oldest endpoint on a timeline whose own dates conflicted and sunset the Assistants API after about eighteen months, Google retired the Gemini CLI into a successor, and image models have gone dark with weeks of notice. Anthropic handles it better than most — the history is kept rather than quietly edited, and it has published commitments to preserve model weights. Good disclosure does not change the planning horizon: about a year of service and two months of warning, whichever logo is on the invoice.
What to do
Export the usage CSV today. Console → Usage → Export breaks consumption down by API key and model — the fastest way to find the Sonnet 4.5 calls nobody remembers writing: the cron job, the staging worker, the internal tool with one user. Anyone standardising on Claude or Claude Code across several teams will find more than expected.
Grep for temperature before anything else. A 400 on the recommended target, present in nearly every Sonnet 4.5 integration, and in Python it fails at the SDK boundary rather than the API. A ten-minute search that otherwise becomes a production incident on 1 December.
Decide 4.6-bridge versus 5.5-destination deliberately. Sonnet 4.6 buys a near-zero-diff move and about ten weeks of extra runway on a legacy model; Sonnet 5.5 costs real engineering time and buys to late September 2027. For a large integration under time pressure the bridge is defensible; for anything you expect to maintain, it is a second migration booked in advance.
Re-measure cost on your own tokens. A 33% list-price cut is roughly 13% on identical text, before any token-efficiency gains. Developers and anyone comparing coding tools or weighing Claude against ChatGPT on price this quarter are comparing numbers that no longer count the same things, and teams on committed-spend agreements should model drawdown against tokenizer-adjusted volume.
Write the Haiku plan, do not run it. Haiku 4.5 has a floor, not a notice. Know the target and the cost; revisit when a deprecation entry appears rather than when a tracker fires.
Assume 60 days from here on. Any model your product depends on should have a named successor and a tested code path before the notice arrives, because the notice is now the start of a two-month sprint rather than a two-quarter plan.
Frequently asked questions
When exactly does Claude Sonnet 4.5 stop working, and on which platforms?
On Anthropic-operated platforms — the Claude API, Claude Platform on AWS and Claude in Microsoft Foundry — requests to claude-sonnet-4-5-20250929 begin failing after 30 November 2026. Anthropic's wording is that requests to models past the retirement date will fail, so plan for errors rather than a silent downgrade. Partner-operated platforms are a separate schedule: Amazon Bedrock and Google Cloud set their own lifecycle dates, and the docs say status and dates can differ there. That split matters if you run multi-cloud for availability, because the same model ID can be retired on one provider and serving on another. Check the Bedrock and Google Cloud model tables directly rather than assuming the first-party date applies everywhere.
Is 61 days' notice unusual, or is that just how Anthropic does it?
It is how Anthropic does it now, and it did not used to be. The published policy promises at least 60 days for publicly released models, and the last five deprecation announcements came in at 62, 60, 62, 61 and 61 days — an average of 61.2. The five before those came in at 63, 181, 189, 76 and 114 days, an average of 124.6. The commitment has not changed; the practice has moved to sit on top of it. For planning purposes the honest assumption is that any Claude model you depend on gets roughly two months of warning, not two quarters, and that the floor is now also the ceiling.
Should the migration go to Sonnet 5.5, Sonnet 5, or Sonnet 4.6?
Anthropic names Sonnet 5.5, and for most teams that is right — it is the only one of the three in the current lineup, it has the longest retirement floor at 28 September 2027, and its June 2026 knowledge cutoff is 17 months fresher than Sonnet 4.5's. The case for Sonnet 4.6 is narrow but real: it keeps the $3/$15 rate card, keeps the older tokenizer, and still accepts temperature, top_p, top_k and extended thinking, so a large legacy integration can move with close to no code change. The cost of that choice is that Sonnet 4.6 is already labelled legacy with a floor of 17 February 2027, which buys about two and a half months past Sonnet 4.5's own retirement. Sonnet 5 sits in between — current pricing, legacy label, 30 June 2027 floor. Treat 4.6 as a bridge you will cross again within a quarter, not a destination.
What is the tokenizer effect, and does it really cancel most of the price cut?
Anthropic's pricing page states that Claude 4.7 and later models use a newer tokenizer that produces approximately 30% more tokens for the same text, and that Sonnet 4.6 and earlier use the previous one. Sonnet 4.5 is on the old tokenizer; Sonnet 5.5 is on the new one. So a document that metered as 1,000,000 tokens on Sonnet 4.5 meters as roughly 1,300,000 on Sonnet 5.5. Input: $3.00 becomes 1.3 × $2.00 = $2.60, about 13% less. Output: $15.00 becomes 1.3 × $10.00 = $13.00, also about 13% less. Cache reads land in the same place — $0.30 against 1.3 × $0.20 = $0.26. The list-price cut is 33%; the effective cut on identical text is closer to 13%. Anthropic is explicit that the exact increase depends on content and workload shape, so treat 13% as the direction of travel and measure your own corpus before committing a number to a budget.
Is Claude Haiku 4.5 being retired on 15 October 2026?
No, and this is worth getting right because several third-party lifecycle trackers have it wrong. The deprecations table lists claude-haiku-4-5-20251001 as Active, with Deprecated marked N/A and a tentative retirement date of 'Not sooner than October 15, 2026.' That is a floor on the earliest possible retirement, not a scheduled end-of-life, and it exists because the commitment was written as one year from the model's October 2025 release. No deprecation notice has been issued. Because Anthropic promises at least 60 days' notice, Haiku 4.5 cannot retire on 15 October even in principle — the soonest date available if a notice went out today would be around 1 December. Auto-generated migration tickets citing 15 October are reading a floor as a deadline. The genuine signal is subtler: Haiku 4.5 is the only model in the current lineup whose floor is inside the next twelve months, so it is the next plausible candidate for a notice.
What breaks if the model string is simply swapped from Sonnet 4.5 to Sonnet 5.5?
Two things will break before any of the model-specific changes do. First, sampling parameters: temperature, top_p and top_k are deprecated from Claude Opus 4.7 onward and return a 400 error when set to a non-default value, and the Python SDK from v1.0 removed them outright so passing them raises a TypeError. Almost every integration written against Sonnet 4.5 sets temperature. Second, extended thinking — the thinking.type: 'enabled' plus budget_tokens mode that Sonnet 4.5 used — is not accepted on models after Sonnet 4.6. Beyond those, Sonnet 5.5 brings five further documented breaking changes and one that returns HTTP 200 while changing what users see. Budget engineering time, not just a config change.
Sources
- Claude Platform Docs — Model deprecations (primary: the 2026-09-30 Sonnet 4.5 deprecation entry, 30 November 2026 retirement, claude-sonnet-5-5 as recommended replacement, the full deprecation history, the 60-day notification commitment, model status table, temperature/top_p/top_k parameter deprecation)
- Claude Platform Docs — Pricing (primary: Sonnet 4.5 at $3/$15, Sonnet 4.6 at $3/$15, Sonnet 5.5 at $2/$10, cache-read rates, and the note that Claude 4.7-and-later models use a tokenizer producing ~30% more tokens for the same text)
- Claude Platform Docs — Models overview (primary: Haiku 4.5 listed in the current lineup with a retirement floor of 15 October 2026 and a February 2025 knowledge cutoff; Sonnet 4.6 and Sonnet 5 listed as legacy; extended thinking not accepted on models after Sonnet 4.6)
- Anthropic Research — Commitments on Model Deprecation and Preservation (primary: weight-preservation commitment and the stated downsides of retirement)
- Wikipedia — Claude (language model) (independent: Claude Sonnet 4.5 public release on 29 September 2025, matching the 20250929 snapshot ID)
- OpenRouter — Claude Sonnet 4.5 model page (independent corroboration: 29 September 2025 availability, $3/$15 pricing, 200K standard context)
- endoflife.date — Anthropic Claude lifecycle tracker (independent: third-party trackers rendering Haiku 4.5's 'not sooner than' floor as a hard end-of-life date)
Related tool reviews
Questions or corrections? Email Pick Right. Want the full list? See all news.