Claude Sonnet 4.5 retirement migration cost calculator

What your Sonnet 4.5 traffic costs on Sonnet 5.5 rates, with tokenizer, cache, batch and thinking-token effects

By Michael Lip, claudflow.com. Updated October 28, 2026.

Short answer: Claude Sonnet 4.5 retires from the Claude API on November 30, 2026. Sonnet 5.5 is the replacement. At list prices input drops from $3 to $2 per million tokens and output from $15 to $10, but the Sonnet 5 tokenizer produces about 30% more tokens for the same text.

Updated October 28, 2026

Anthropic announced on September 30, 2026 that Claude Sonnet 4.5 (claude-sonnet-4-5-20250929) retires from the Claude API on November 30, 2026. The recommended replacement is Claude Sonnet 5.5 (claude-sonnet-5-5). Enter your monthly token volumes. The calculator applies both rate cards, the Sonnet 5 tokenizer factor, your cache mix, the Batch API discount, the US-only inference premium on Sonnet 5.5, and Sonnet 5.5 thinking tokens, then shows the monthly and annual difference.

Days remaining until retirement: -

Uncached input, million tokens per month

Output, million tokens per month

Cache read, million tokens per month

Cache write 5m, million tokens per month

Cache write 1h, million tokens per month

Sonnet 5.5 thinking tokens, million per month (billed as output)

Tokenizer factor on Sonnet 5.5 (about 1.30 for text, depends on content)

Batch API share, percent of token volume

Endpoint

Monthly cost at Sonnet 4.5 rates-

Monthly cost at Sonnet 5.5 rates-

Monthly difference after migration-

Change-

Annualized difference-

Enter your monthly token volumes to see the cost difference between Sonnet 4.5 and Sonnet 5.5 rates.

If the thinking input is 0, the result does not include thinking tokens. Sonnet 5.5 runs adaptive thinking by default, so enter your estimate of its thinking tokens. If you used extended thinking on Sonnet 4.5, those thinking tokens are already in your billed output. The output field carries them to Sonnet 5.5 with the tokenizer factor, so enter in the thinking field only the thinking you expect above that amount.

How the calculation works

token counts (MTok) * 1,000,000 = tokens Sonnet 4.5 rates (USD per MTok): input 3, output 15, cache read 0.30, cache write 5m 3.75, cache write 1h 6 Sonnet 5.5 rates (USD per MTok): input 2, output 10, cache read 0.10, cache write 5m 2.50, cache write 1h 4 Sonnet 4.5 cost = sum of (token count * Sonnet 4.5 rate) Sonnet 5.5 cost = (sum of (token count * tokenizer factor * Sonnet 5.5 rate) + thinking * 10) * geo geo = 1.1 for US-only inference on Sonnet 5.5, else 1 (Sonnet 4.5 has no inference_geo) Batch API bills 50% of standard prices on the batched share of both costs month difference = Sonnet 4.5 cost - Sonnet 5.5 cost

The tokenizer factor applies on Sonnet 5.5 only. Anthropic's pricing page says the newer tokenizer produces approximately 30% more tokens for the same text. Token counts you measured on Sonnet 4.5 grow on Sonnet 5.5.
Sonnet 5.5 runs adaptive thinking by default on requests with no thinking field. Sonnet 4.5 ran without thinking by default. Thinking tokens are billed as output tokens, and Anthropic's migration guide says max_tokens covers thinking plus text. The calculator bills the thinking input at the Sonnet 5.5 output rate only and does not apply the tokenizer factor to it.
The Batch API option halves the batched share of every line item. Anthropic's batch documentation says "All usage is charged at 50% of the standard API prices."
US-only inference (the inference_geo parameter) costs 1.1x on all Sonnet 5.5 token pricing categories. Sonnet 4.5 does not accept inference_geo (requests that include it return a 400 error), so the Sonnet 4.5 cost always uses standard pricing. Global routing uses standard pricing.
Cache write prices are the same multiples of base input on both models (1.25x for 5-minute writes, 2x for 1-hour writes), so a cache write token changes cost by the same ratio as an uncached input token. Cache reads change most: 0.1x base input on Sonnet 4.5 ($0.30) versus 0.05x on Sonnet 5.5 ($0.10), so cache-heavy traffic benefits most.
Token counts are estimates. Count tokens with the token counting API on each model ID before you re-budget.
The calculator is for text tokens. Sonnet 5.5 uses a high-resolution image tier, and Anthropic's migration guide says a 2000×1500 image costs about 2.5 times as many tokens, so do not apply the text tokenizer factor to image tokens.

This tool does not execute LLM calls. All computation happens locally in your browser.

Reference rates

Verified 2026-10-20 against the Anthropic pricing page (https://platform.claude.com/docs/en/about-claude/pricing). Units are USD per million tokens.

Standard prices, USD per million tokens
Token type	Sonnet 4.5	Sonnet 5.5
Input	$3.00	$2.00
Output	$15.00	$10.00
Cache read	$0.30	$0.10
Cache write, 5 minutes (1.25x input)	$3.75	$2.50
Cache write, 1 hour (2x input)	$6.00	$4.00
Batch API (50% of standard)	$1.50 / $7.50	$1.00 / $5.00

Retirement facts

Anthropic notified developers on September 30, 2026 that claude-sonnet-4-5-20250929 retires from the Claude API on November 30, 2026.
After the retirement date, requests to the old model ID fail. The migration is mandatory for Claude API traffic.
The November 30, 2026 date applies to the Claude API, Claude Platform on AWS and Microsoft Foundry. Amazon Bedrock and Google Cloud set their own retirement schedules, and this page uses Claude API prices.
The September 30, 2026 notice covers only Claude Sonnet 4.5. Other models have their own dates on the model deprecations page.

Migration checklist from Anthropic

Condensed from the Sonnet 5.5 migration guide (https://platform.claude.com/docs/en/models/sonnet-5-5/migration-guide). Follow the links there for the full text.

Every starting model

Change the model ID to claude-sonnet-5-5, which has no date suffix.
Pass thinking blocks back unchanged, and read content blocks by type.
To run without up-front thinking, send the lowest thinking setting at high effort or below.
Replace forced tool use with auto and strict tools.
Keep conversations append-only.
On the Claude API and Google Cloud, move computer use to the toolset, without the fine-grained-tool-streaming-2025-05-14 beta header.
Pair the advisor tool with a supported advisor, and expect encrypted advice.
Read text between tool calls from thinking blocks.
Handle refusals, and configure fallback.
Re-run your effort sweep and re-baseline cost.

Claude Sonnet 4.6 or earlier

Expect thinking on requests with no thinking field, and revisit max_tokens.
Replace thinking budgets with an effort level.
Remove non-default temperature, top_p, and top_k values.
If you show thinking text, set display: "summarized".
Recount tokens, and re-budget image tokens.

Claude Sonnet 4.5 or earlier

Replace assistant prefills. Sonnet 5.5 rejects a prefilled last assistant turn with a 400 error.
Parse tool call input with a standard JSON parser.
On Amazon Bedrock, move computer use from computer_20250124 to computer_20251124.
Set output_config.effort explicitly. Sonnet 4.5 has no effort parameter.
Remove any context-window beta header.
Remove interleaved-thinking-2025-05-14, and replace fine-grained-tool-streaming-2025-05-14 with eager_input_streaming.
Move output_format to output_config.format.

Frequently asked questions

When does Claude Sonnet 4.5 retire from the Claude API?

November 30, 2026. Anthropic notified developers on September 30, 2026.

Which model replaces Claude Sonnet 4.5?

Anthropic recommends claude-sonnet-5-5 as the replacement.

Does Claude Sonnet 5.5 cost less than Sonnet 4.5?

At list prices it costs less per token. Input is $2 per million tokens versus $3, output is $10 versus $15, and cache reads drop from $0.30 to $0.10. The same text produces about 30% more tokens on the Sonnet 5 tokenizer, so enter the tokenizer factor above before you compare totals. Sonnet 5.5 also runs adaptive thinking by default, and thinking tokens are billed as output, so enter your expected thinking tokens too. Cache-heavy workloads still see the largest relative drop.

Do I have to migrate before the retirement date?

On the Claude API, yes. Requests to claude-sonnet-4-5-20250929 fail once the model retires on November 30, 2026. Anthropic's notice gives no cutoff hour, so finish the move before that day. Migrating earlier moves traffic to the lower Sonnet 5.5 per-token prices sooner and leaves time to re-run evaluations and recount tokens before the cutoff.

How do assistant prefills behave after migration?

Sonnet 5.5 rejects a prefilled last assistant turn with a 400 error that reads "This model does not support assistant message prefill." Replace each prefill by its purpose: structured outputs for output format, system prompt instructions for preambles, clearer user turns for unwanted refusals, user-message continuations for interrupted responses, and the user turn for context reminders.