Claude Sonnet 4.5 Retires on November 30, 2026: What Moving to Claude Sonnet 5.5 Does to the Bill

On November 30, 2026, Anthropic retires Claude Sonnet 4.5 from its API. The model was released on September 29, 2025, deprecated on September 30, 2026, and has two months of life left. This post reads, as they stood on October 11, 2026, the four Anthropic pages that matter for the move: the deprecations page, the model page of Sonnet 4.5, the migration guide of the recommended replacement and the general pricing page. It then puts a price on the move at two fixed workloads and lists what else Anthropic and other makers charge for at least the same context window. It compares prices only. Nothing on these pages measures which model answers better, and this site does not either.

The retirement is also on this site’s deprecations calendar, with Anthropic’s own words.

The date and the replacement, as Anthropic prints them

Anthropic’s deprecations page has an entry headed “2026-09-30: Claude Sonnet 4.5 model”, which reads: “On September 30, 2026, Anthropic notified developers using Claude Sonnet 4.5 of its upcoming retirement on the Claude API.” Its table has one row:

• November 30, 2026 | claude-sonnet-4-5-20250929 | claude-sonnet-5-5 |

The same page’s status table says the same thing in another form: claude-sonnet-4-5-20250929 | Deprecated | September 30, 2026 | November 30, 2026 |. Both rows are identical to what this site first recorded on October 2.

The model page of Claude Sonnet 4.5 opens with a banner: “Claude Sonnet 4.5 is deprecated: it is still functional but no longer recommended, and retires on November 30, 2026. Recommended replacement: Claude Sonnet 5.5.” Its availability block prints “Status Deprecated”, “Deprecated September 30, 2026” and “Retires November 30, 2026”. It lists the alias claude-sonnet-4-5 next to the dated ID, so code that never pinned the snapshot is affected too.

The deprecations page covers the Claude API. The model page also lists Amazon Bedrock, Google Cloud, Microsoft Foundry and Claude Platform on AWS as platforms; their schedules are for those platforms to publish, and this post does not cover them.

What the two models cost

Every price below was read on October 11, 2026 from Anthropic’s pricing page, per million tokens, in US dollars. Sonnet 4.5 now sits under “Additional models”; its row reads Claude Sonnet 4.5 | $3 / MTok | $15 / MTok | $3.75 / MTok | $6 / MTok | $0.30 / MTok, and the row of Claude Sonnet 5.5 reads $2 / MTok | $10 / MTok | $2.50 / MTok | $4 / MTok | $0.10 / MTok, in the order input, output, 5-minute cache write, 1-hour cache write, cache read. The Batch table, under “The Batch API allows asynchronous processing of large volumes of requests with a 50% discount on both input and output tokens”, prints $1.50 / $7.50 for Sonnet 4.5 and $1 / $5 for Sonnet 5.5.

Price per 1M tokens	Claude Sonnet 4.5	Claude Sonnet 5.5
Input	$3.00	$2.00
Output	$15.00	$10.00
Cache write, 5 minutes	$3.75	$2.50
Cache write, 1 hour	$6.00	$4.00
Cache read	$0.30	$0.10
Batch input / output	$1.50 / $7.50	$1.00 / $5.00
Context window	200,000	1,000,000
Long-context tier	none	none

Neither model has a long-context tier. Sonnet 4.5 stops at “Context window 200K tokens”, so no prompt can grow into one. For Sonnet 5.5, the pricing page says that “Claude 4.6 and later models (except Claude Haiku 5.5) and Claude Mythos Preview include the full 1M token context window at standard pricing”, and adds: “A 900k-token request is billed at the same per-token rate as a 9k-token request.”

Every line is lower on Sonnet 5.5: input, output and cache writes by a third, and the cache read by two thirds. That last figure is recent. On October 7, 2026 Anthropic’s release notes said: “We’ve lowered the price of prompt cache reads on Claude Sonnet 5.5 from $0.20 USD to $0.10 USD per million tokens: 0.05x the base input price instead of 0.1x.” The pricing page now carries the exception in its caching table: “Cache read (hit) | 0.1x base input price (0.025x on Claude Fable 5.1 and Claude Mythos 5.1; 0.05x on Claude Opus 5.5 and Claude Sonnet 5.5)”. So a Sonnet 4.5 cache hit costs 10% of its input price and a Sonnet 5.5 hit costs 5%.

The bill at two fixed workloads

Two workloads, fixed so the comparison is the same for every model: 10M input tokens and 2M output tokens a month with no cache, and the same volume with 80% of the input read from the cache. The arithmetic is the one the cost calculator uses: tokens divided by a million, times the price per million, input, cached input and output added. Cache writes are not counted, as the calculator does not count them; filling the cache costs $3.75 per million tokens written on Sonnet 4.5 and $2.50 on Sonnet 5.5 for the 5-minute cache, so it does not change which model is cheaper.

Workload	Claude Sonnet 4.5	Claude Sonnet 5.5	Change
10M in, 2M out, no cache	$60.00	$40.00	$20.00 less
Same, 80% of input from cache	$38.40	$24.80	$13.60 less
Same as first, Batch API	$30.00	$20.00	$10.00 less

The cached line splits like this. On Sonnet 4.5, 2M fresh input tokens cost $6.00, 8M cache reads cost $2.40 and 2M output tokens cost $30.00. On Sonnet 5.5 the same three parts cost $4.00, $0.80 and $20.00. Output is most of both bills, which is why a cheaper cache read moves the total less than its two-thirds cut suggests.

The October 7 cut is worth exactly $0.80 a month at this workload. At the old $0.20 cache read, the cached Sonnet 5.5 bill was $25.60, and the saving over Sonnet 4.5 was $12.80; it is now $13.60. The cut helps most where cache reads are a large share of the bill: long system prompts or documents reused across many short answers.

Where the saving can shrink: tokens, not prices

The table assumes the same token counts on both models. Anthropic’s migration guide for Sonnet 5.5 says that is not the case: “Claude Sonnet 5.5 uses Claude Sonnet 5’s tokenizer. Against Claude Sonnet 4.6, Claude Sonnet 4.5, and Claude Haiku 4.5, the same text produces about 30% more tokens, depending on the content.”

If the increase were exactly 30% on both input and output, the workload measured on Sonnet 4.5 as 10M in and 2M out becomes 13M and 2.6M on Sonnet 5.5. The uncached bill on Sonnet 5.5 is then $52.00, still $8.00 under the $60.00 of Sonnet 4.5; the cached one is $32.24 against $38.40. The prices give some room: Sonnet 5.5 costs a third less per token, so the same text would have to produce 50% more tokens on every line before the uncached bills met. Anthropic gives about 30% and says it depends on the content, so the only way to know a given workload’s figure is to count it: the guide asks to recount with token counting and to revisit max_tokens and cost.

The second source of extra tokens is thinking. The guide says that on Sonnet 5.5 “a request with no thinking field runs with adaptive thinking”, where on Sonnet 4.5 the same request “ran without thinking”, and that “thinking tokens are billed as output tokens”. At the uncached workload, with token counts unchanged, Sonnet 5.5 would need 4M output tokens a month instead of 2M, twice as many, to reach Sonnet 4.5’s $60.00. With the 30% tokenizer increase already counted, the room is 0.8M extra output tokens. To keep running without up-front thinking, the guide says to send thinking: {"type": "between_tools"}; sending {"type": "disabled"} returns an error on this model.

What else changes in the code, where Anthropic says so

The guide’s checklist for code coming from “Claude Sonnet 4.5 or earlier” lists the steps. This post repeats the ones that can break a request:

• Prefill. “Claude Sonnet 5.5 rejects a prefilled last assistant turn with a 400 error”. Sonnet 4.5 accepted one.
• Thinking budgets and sampling. budget_tokens returns a 400 error; the guide says to set an effort level instead: “There’s no fixed mapping from a budget to an effort level”. Non-default temperature, top_p and top_k also return a 400 error.
• Forced tool use. A tool_choice of type any or tool returns a 400 error; the guide gives auto with strict tools instead.
• Computer use and beta headers. computer_20250124 is not accepted on any platform; interleaved-thinking-2025-05-14 should be removed and output_format moved to output_config.format.
• Caching minimum. “The minimum cacheable prompt is 512 tokens, down from 1,024 on Claude Sonnet 5, Claude Sonnet 4.6, and Claude Sonnet 4.5.” Prompts between those sizes can be cached for the first time.

Alternatives from Anthropic, by price only

Every current Anthropic model has at least Sonnet 4.5’s 200,000-token window, so all of them qualify. These are the ones that cost no more than Sonnet 4.5 at the same two workloads, cheapest first, from the pricing page read on October 11, 2026:

Model	Input / output per 1M	Cache read	No cache	80% cached	Context
Claude Haiku 5.5	$0.10 / $0.50	$0.01	$2.00	$1.28	1,000,000
Claude Haiku 4.5	$1.00 / $5.00	$0.10	$20.00	$12.80	200,000
Claude Sonnet 5.5	$2.00 / $10.00	$0.10	$40.00	$24.80	1,000,000
Claude Sonnet 5	$2.00 / $10.00	$0.20	$40.00	$25.60	1,000,000
Claude Sonnet 4.6	$3.00 / $15.00	$0.30	$60.00	$38.40	1,000,000

Three notes on that table. Claude Haiku 5.5’s prices hold only for prompts up to 100,000 tokens; above that, Anthropic’s page prints $0.50 input and $2.50 output, and a Sonnet 4.5 prompt can reach 200,000. Claude Haiku 4.5 is “Active (legacy)”, with a retirement “Not sooner than October 15, 2026” on the deprecations page, four days after this reading. Claude Sonnet 4.6 costs exactly what Sonnet 4.5 costs, with a 1M window and a retirement “Not sooner than February 17, 2027”; moving there changes the date, not the bill. Everything else Anthropic sells, from Claude Opus 5.5 at $80.00 a month uncached upward, costs more than Sonnet 4.5.

Other makers in the same price band

A Sonnet 4.5 workload today costs $60.00 uncached; Sonnet 5.5 brings it to $40.00. These are the current models from other makers in this site’s dataset with at least 200,000 tokens of context whose uncached bill falls between those two figures, cheapest first:

Model	Input / output per 1M	Cache read	No cache	80% cached	Verified
GPT-6 Sol	$2.00 / $10.00	$0.20	$40.00	$25.60	October 5, 2026
Qwen3.7-Max	$2.50 / $7.50	$0.50	$40.00	$24.00	October 11, 2026
GPT-5.6 Terra	$2.00 / $12.00	$0.20	$44.00	$29.60	October 5, 2026
Gemini 3.1 Pro Preview	$2.00 / $12.00	$0.20	$44.00	$29.60	October 5, 2026

The OpenAI and Google models charge more for long prompts: above 272,000 input tokens for the two GPT rows and above 200,000 for Gemini 3.1 Pro Preview. A prompt that fits in Sonnet 4.5’s 200,000-token window stays below both thresholds, so the first band is the one that applies to anything that runs on Sonnet 4.5 today. Qwen3.7-Max is Alibaba’s Singapore (“Scope: International”) price. At the cached workload it is the only model in this table that undercuts Sonnet 5.5, by $0.80, because its output price is lower; at the uncached one it ties.

Cheaper models with long context exist at every maker; the pricing table and the long-prompt ranking list them. The tokenizer caveat above applies across makers too: the same text is not the same number of tokens on two models, and none of these tables can say how many tokens a given workload takes elsewhere.

How this was read

The date, the replacement and the migration notes come from Anthropic’s deprecations page, the model page of Claude Sonnet 4.5 and the Claude Sonnet 5.5 migration guide, downloaded on October 11, 2026. Every Anthropic price comes from the pricing page read the same day; the other makers’ prices come from this site’s dataset, where every row links to the page it was read on, with the date shown. Prices are standard list prices per million tokens in US dollars, without data residency or regional endpoint premiums. The retirement stays on the deprecations calendar until November 30, 2026.

Frequently asked questions

When is Claude Sonnet 4.5 retired, and what replaces it?

Anthropic retires claude-sonnet-4-5-20250929 on November 30, 2026, and its deprecations page names claude-sonnet-5-5 as the recommended replacement, as read on October 11, 2026. The model was deprecated on September 30, 2026.

Is Claude Sonnet 5.5 cheaper than Claude Sonnet 4.5?

Per token, yes: $2 input and $10 output per million against $3 and $15, so 10M input and 2M output tokens a month cost $40.00 instead of $60.00. With 80% of the input read from cache the figures are $24.80 and $38.40.

Does the new tokenizer cancel the saving?

Anthropic says the same text produces about 30% more tokens on Claude Sonnet 5.5 than on Claude Sonnet 4.5. At exactly 30% more, the uncached workload costs $52.00 on Sonnet 5.5, still $8.00 less than $60.00.

What did the October 7 cache price cut change for this move?

Anthropic lowered Sonnet 5.5 cache reads from $0.20 to $0.10 per million tokens on October 7, 2026. At the cached workload that takes the Sonnet 5.5 bill from $25.60 to $24.80.