Claude Sonnet 5.5 Cache Reads Fall to $0.10: The Bill Math

On October 7, 2026 Anthropic halved one price of Claude Sonnet 5.5: reading a prompt from the cache now costs $0.10 per million tokens instead of $0.20. Nothing else moved. This post checks the change against Anthropic’s own pages, read again on October 13, 2026, puts it next to the cache prices of the models around it, and works out what it does to the monthly bill of one cached workload at three hit rates. The release notes give no reason for the change, and this post does not guess one.

What Anthropic changed, in its own words

The platform release notes entry for October 7, 2026 reads: “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. Cache writes and all other prices are unchanged.”

Anthropic’s pricing page, read on October 13, 2026, prints the row as Claude Sonnet 5.5 | The best combination of speed and intelligence | $2 / MTok | $10 / MTok | $2.50 / MTok | $4 / MTok | $0.10 / MTok (input, output, 5-minute cache write, 1-hour cache write, cache read). Its caching table now names the model in the exception: “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)”. This site recorded the change on its price-changes page as a decrease by the provider, effective October 7, 2026, after reading the new figure on October 11. All five figures on the page match the dataset row, verified on October 12.

Claude Sonnet 5.5, per 1M tokens	Until October 6, 2026	From October 7, 2026
Input	$2.00	$2.00
Output	$10.00	$10.00
Cache write, 5 minutes	$2.50	$2.50
Cache write, 1 hour	$4.00	$4.00
Cache read	$0.20	$0.10

Next to the other Claude models

The models a Sonnet 5.5 user is most likely to weigh it against, with the prices on Anthropic’s page on October 13, 2026:

Model	Status in this dataset	Input	Cache read	Cache read as share of input
Claude Haiku 5.5	current	$0.10	$0.01	0.1x
Claude Sonnet 5.5	current	$2.00	$0.10	0.05x
Claude Opus 5.5	current	$4.00	$0.20	0.05x
Claude Sonnet 4.6	current	$3.00	$0.30	0.1x
Claude Sonnet 4.5	deprecated, retires November 30, 2026	$3.00	$0.30	0.1x

Haiku 5.5’s figures are its list for prompts up to 100,000 tokens; above that line every one of its prices is five times higher, as our post on its threshold works through. Claude Sonnet 5, which Anthropic now lists under “Additional models”, still prints $0.20 for cache reads with the same $2, $10, $2.50 and $4 as Sonnet 5.5, so since October 7 the two differ on that one line only.

Sonnet 5.5 came out after our cache price survey of September 16, which found a cache hit at 10% of the input price to be the norm. It launched at that norm, $0.20 on $2, and since October 7 sits below it, with Opus 5.5, Fable 5.1 and Mythos 5.1.

What the cache-write rule adds

A cache read is only half of the arithmetic. The pricing page’s caching table lists “5-minute cache write | 1.25x base input price | Cache valid for 5 minutes” and “1-hour cache write | 2x base input price | Cache valid for 1 hour”, and explains when each is charged: “Cache write tokens are charged when content is first stored. Cache read tokens are charged when a subsequent request retrieves the cached content.” The October 7 change touched neither write price, so on Sonnet 5.5 a 5-minute write still costs $2.50 per million tokens, 25 times the new read price.

One workload, three hit rates

Take a chatbot or a retrieval setup with a long system prompt that repeats on every request: 100 million input tokens and 20 million output tokens a month on Sonnet 5.5. The hit rate is the share of the input tokens read from the cache. To keep the count simple and on the expensive side, every input token that is not a hit is treated as a 5-minute cache write at $2.50, which is what a request pays when it stores the prefix that later requests read. Output costs $200.00 a month in every case. Without any caching the same month costs $400.00.

Cache hits	Cache writes	Reads before	Reads after	Bill before	Bill after	Saving
50%	$125.00	$10.00	$5.00	$335.00	$330.00	$5.00 (1.5%)
80%	$50.00	$16.00	$8.00	$266.00	$258.00	$8.00 (3.0%)
95%	$12.50	$19.00	$9.50	$231.50	$222.00	$9.50 (4.1%)

With 1-hour writes instead, at $4, the bills are higher, $405.00, $288.00 and $229.50 after the change at the three rates, but the saving is identical: the change lowered reads, and reads are the same in both cases.

Where the bill changes, and by how much

The saving follows a single rule: $0.10 for every million tokens read from cache. On 100 million input tokens it cannot pass $10 a month, whatever the hit rate. What grows with the hit rate is its share of the bill, because cache writes shrink as hits rise: 1.5% at 50%, 3.0% at 80% and 4.1% at 95%. Under 80% hits the saving stays under 3% of the bill; at 50% the month spends $125.00 writing to the cache and saves $5.00 on reading it.

For the same workload the hit rate matters far more than the price. Going from 80% to 95% hits lowers the post-change bill from $258.00 to $222.00, $36.00, nearly four times what the price cut saves at 95%. A team that wants a smaller Sonnet 5.5 bill gets more from keeping the cached prefix identical between requests, and inside the cache’s lifetime, than from the October 7 change.

The rule scales. A workload with ten times the volume, 1 billion input tokens a month at 95% hits, saves $95.00; the 4.1% holds only if output grows in step.

Against the other models, same workload

The same 100 million input and 20 million output tokens, with the same treatment of misses as 5-minute writes, at each model’s current list:

Model	50% hits	80% hits	95% hits
Claude Haiku 5.5 (prompts up to 100,000 tokens)	$16.75	$13.30	$11.58
Claude Sonnet 5.5	$330.00	$258.00	$222.00
Claude Sonnet 4.6 and 4.5	$502.50	$399.00	$347.25
Claude Opus 5.5	$660.00	$516.00	$444.00

Output dominates every row: on Sonnet 5.5 it is $200.00 of the $222.00 at 95% hits. That is why a cheaper cache read barely reorders the table. Sonnet 5.5 was already cheaper than Sonnet 4.6 on every line of its price list; the cut widens the gap by the $9.50 shown above at 95% hits. For anyone still on Sonnet 4.5, which Anthropic retires on November 30, 2026, the move to Sonnet 5.5 that Anthropic recommends takes this bill from $347.25 to $222.00 at 95% hits. The comparison is of prices only; nothing here measures how the models answer.

How this was read

Anthropic’s figures and sentences come from its pricing page and its platform release notes, downloaded on October 13, 2026, and match this site’s dataset rows for the five models, verified on October 12, 2026, and the price-change entry recorded on October 11. Prices are standard list prices in US dollars, without Batch discounts or data residency premiums, which the pricing page says stack with the cache multipliers. All of Anthropic’s current prices are on the Anthropic pricing page of this site.

Frequently asked questions

What did Anthropic change on Claude Sonnet 5.5 on October 7, 2026?

Only the cache-read price, from $0.20 to $0.10 per million tokens, 0.05x the $2 input price instead of 0.1x. Input ($2), output ($10) and cache writes ($2.50 for 5 minutes, $4 for 1 hour) did not change, as the pricing page still showed on October 13, 2026.

How much does the cut save on a monthly bill?

On 100M input and 20M output tokens a month, $5 at 50% cache hits, $8 at 80% and $9.50 at 95%, or 1.5%, 3.0% and 4.1% of the bill. The saving is $0.10 for every million tokens read from cache, so it can never pass $10 on 100M input tokens.

Is Claude Sonnet 5.5 now cheaper than Claude Sonnet 5?

On cache reads only: Claude Sonnet 5 still lists $0.20, and every other price of the two models is the same, $2 input, $10 output, $2.50 and $4 cache writes.