Claude Sonnet 4.5 retires on 30 November: what the migration really costs
Sonnet 4.5 shuts down on 30 November 2026. Its replacement costs 33% less per token but counts about 30% more tokens, so the real saving is about 13%.
Anthropic deprecated Claude Sonnet 4.5 on 30 September 2026 and will retire it on 30 November. The named replacement, Claude Sonnet 5.5, costs a third less per token. Your bill will not fall by a third, because Sonnet 5.5 counts your text differently.
The dates
- 30 September 2026: Claude Sonnet 4.5 (
claude-sonnet-4-5-20250929) deprecated. - 30 November 2026: retired. Requests stop working.
- Replacement: Claude Sonnet 5.5, released 28 September 2026.
That is a 60-day window, which is short for anything with an evaluation suite and a release process behind it.
The headline saving
| Per 1M tokens | Sonnet 4.5 | Sonnet 5.5 | Change |
|---|---|---|---|
| Input | $3.00 | $2.00 | 33% lower |
| Output | $15.00 | $10.00 | 33% lower |
| Cache read | $0.30 | $0.20 | 33% lower |
| 5-minute cache write | $3.75 | $2.50 | 33% lower |
Every rate is a third lower. Read no further and you would plan a 33 percent saving.
The tokenizer catch
Anthropic's documentation says plainly: "Claude 4.7 and later models and Claude Mythos Preview use a newer tokenizer. The same input text produces approximately 30 percent more tokens than on earlier models." It also says that Claude Sonnet 4.6 and earlier use the previous tokenizer.
Sonnet 4.5 is on the old tokenizer. Sonnet 5.5 is on the new one. So the same prompt, unchanged, becomes about 30 percent more tokens the moment you switch, and billing follows the new count.
Anthropic's own migration advice is to recount: do not reuse token counts measured on the older model, and count your prompts against the model ID you plan to use.
The real change
Put the two together. A prompt that counted 1,000 tokens on Sonnet 4.5 counts about 1,300 on Sonnet 5.5. The effective price of the same text is the new rate multiplied by the new count:
| For the same text | Sonnet 4.5 | Sonnet 5.5 | Real change |
|---|---|---|---|
| Input, per 1M old tokens | $3.00 | $2.60 | 13% lower |
| Output, per 1M old tokens | $15.00 | $13.00 | 13% lower |
About 13 percent, not 33. Still a saving, and still worth making, but a budget built on the headline figure will be wrong by a factor of two and a half.
Two things could move that number in your favour. Anthropic says the exact increase depends on the content, so your own measurement may differ from 30 percent. And Anthropic claims Sonnet 5.5 produces fewer output tokens for the same work than its predecessor, which would reduce the billed output volume even as the count per word rises. Both are reasons to measure rather than assume.
Context fit changes too. A prompt that fitted in a context window with little room to spare now needs about 30 percent more space, so re-check anything near the limit.
Migration checklist
- Recount your real prompts against
claude-sonnet-5-5, not against your old numbers. Anthropic's token counting endpoint is free. - Re-check context fit for the longest prompts you send.
- Re-measure output length. The output side is where the saving is largest and where model behaviour changed.
- Review cache settings. The minimum cacheable prompt is 512 tokens on Sonnet 5.5 against 1,024 on Sonnet 4.5, so shorter prefixes can now be cached.
- Finish before 30 November. Retirement is a hard stop, not a deprecation warning.
This site now counts Claude 4.7 and later with a denser projection than older Claude models for exactly this reason. Paste a real prompt into the token calculator and compare, or see Sonnet 5.5 against Opus 5.5.
Evidence ledger
Reviewed: 2 October 2026 · Anthropic's documentation remains authoritative.