The Cache Beats the Model: What Really Drives an AI Bill

Anthropic released Claude Fable 5.1 on 1 September 2026. The list price stayed where it was: 10 US dollars per million input tokens, 50 US dollars per million output tokens. What came down is a number that hardly appears in the usual price comparisons.

It is the price of a cache hit, and it fell from 1.00 to 0.25 US dollars per million tokens. In a worked monthly bill for a back-office workflow, that single line accounts for the entire price difference between Fable 5 and Fable 5.1: 1,512 US dollars a month. And it is not even the biggest lever.

Key points: Anthropic released Claude Fable 5.1 and Claude Mythos 5.1 on 1 September 2026. Input and output still cost 10 and 50 US dollars per million tokens, while the price of a cache hit fell 75 percent to 0.25 US dollars. Anthropic puts the saving at around 25 percent for typical workloads and up to about 45 percent for highly agentic ones. A worked monthly bill for 4,200 cases with eight model calls each confirms it: 3,830.40 US dollars on Fable 5, 2,318.40 on Fable 5.1, a reduction of 39.5 percent. The bigger lever sits elsewhere. The same workflow without caching costs 21,672 US dollars, nine times as much. The ranking is therefore: architecture before model tier before model version. For mid-sized applications it also matters that Sonnet 5 stays at 2 and 10 US dollars; the increase to 3 and 15 announced for 1 September did not take place. Update of 14 October 2026: Since this article was published, Anthropic has released Opus 5.5, Sonnet 5.5 and Haiku 5.5. On Opus 5.5 and Sonnet 5.5 a cache hit costs five percent of the input price, and for requests up to 100,000 tokens Haiku 5.5 costs 90 percent less per token than Haiku 4.5. Where Haiku 5.5 is up to the task, the version jump and the model tier each do more in the example than caching.

What did Anthropic change on 1 September?

Two models appeared: Claude Fable 5.1 and Claude Mythos 5.1. Anthropic describes them as identical, with Mythos 5.1 carrying less restrictive safeguards and, at release, being open only to vetted organisations in the United States working in cybersecurity or the life sciences. For companies in Germany, Fable 5.1 is therefore the relevant version.

On price, Anthropic states the decisive sentence itself: cache reads now cost 75 percent less, namely 0.25 US dollars per million tokens. Everything else stays the same. How unusual that is only becomes visible in the full price list.

All figures per million tokens, as of 2 September 2026.

The multiplier in the second-to-last column is the remarkable part. On 2 September, a cache hit cost ten percent of the input price across the entire Claude price list, and Fable 5.1 and Mythos 5.1 were the only exceptions at 2.5 percent. Anthropic did not turn a dial that applies to everyone; it introduced a special rule for its most expensive tier. Since 22 September 2026 there has been a further level of five percent in between, today for Opus 5.5 and Sonnet 5.5; more on that in the update at the end.

Commercially that makes sense. Fable is built for long-running agents, and in exactly those workloads the input consists mostly of repetition. Cutting the price of that share cuts the bill for the intended audience without touching the list price that comparisons and negotiations anchor on.

What is a cache hit and why does it rarely show up in comparisons?

A language model processes every request from the beginning. If a workflow sends the same system instruction, the same rules and the same worked examples eight times in a row, those tokens are computed eight times and paid for eight times.

Prompt caching holds the unchanged beginning of a request. The first time it is written, which costs between 1.25 times and twice a normal input. On every further call within the retention period it is only read, and that read is the cache hit. Two retention periods are available: five minutes, with writes at 1.25 times the input price, and one hour, with writes at twice the input price.

Anyone who builds such workflows sees the pattern quickly. The fixed prefix is almost always the largest single item of input, because it carries all the rule knowledge the workflow needs: pricing logic, approval limits, exceptions, sample cases. The actual request next to it is tiny. That is precisely why the price of a cache hit decides the bill, not the price of input.

In price comparisons that number rarely shows up. What gets compared is input and output, because those are the two values every provider quotes.

What does a back-office agent cost per month?

To make the order of magnitude verifiable, here is a fully worked example. It describes a workflow of the kind found in quote or invoice checking.

That yields 15.12 million written, 2,016 million read and 50.4 million fresh input tokens plus 20.16 million output tokens. Multiplied by the price list, the monthly bill looks like this.

All values in US dollars per month.

The volumes apply to Fable 5, Fable 5.1, Opus 5 and Sonnet 5. These models use a new tokenizer that splits text into around 30 percent more tokens than older models do (more on this below). Haiku 4.5 still uses the older one and counts around 23 percent fewer tokens for the same text; for the same text its column would therefore come to around 295 instead of 383.04 US dollars. Twelve write operations per working day is a cautious assumption: every hit refreshes the entry, so continuous operation needs fewer.

Moving from Fable 5 to Fable 5.1 lowers the bill by 1,512 US dollars, or 39.5 percent. That sits at the upper end of the range Anthropic quotes, which fits the setup: a 60,000-token prefix read eight times per case is a strongly agentic pattern.

The reason is visible too. On Fable 5, cache hits made up 52.6 percent of the entire bill, more than any other line item. On Fable 5.1 they account for 21.7 percent. The largest single item has become output, and that is a different optimisation problem.

A second observation from the same table: Opus 5 costs 1,915.20 US dollars, less than Fable 5.1 at 2,318.40, even though Opus 5 received no special cache price. The discount narrows the gap between the tiers, it does not reverse it.

Model costs are only one item here. What setup, integration and operation of an automation cost on top is set out in What does AI automation cost in 2026?.

Which lever is the biggest?

So far this has been about model choice. The far larger item, however, sits in how the workflow is built. Run the same scenario without caching and the prefix is computed fresh on every one of the 33,600 calls.

That settles the ranking of the levers:

1. How the workflow is built decides a factor of 9. Whether the recurring part is cached at all is the single most expensive decision.
2. The model tier decides a factor of 3. Sonnet 5 instead of Fable 5.1 saves 1,552 US dollars a month in this scenario, where the task allows it.
3. The model version decides a factor of 1.65. The jump from Fable 5 to Fable 5.1 is the smallest of the three.

Anyone discussing AI costs while talking only about models is therefore debating the smallest of the three items. This holds regardless of provider: OpenAI, Google and Anthropic bill caching differently, but all three bill it far cheaper than fresh input.

Since Haiku 5.5 this ranking no longer holds everywhere: where Haiku 5.5 is up to the task, the version jump and the model tier each do more in the example than caching (see the update).

How the minimum length makes a cheap model more expensive

There is one condition that does not appear in the price list and still decides the bill: the minimum length. A text has to reach a certain size before a cache entry is created at all. That threshold differs by model.

The table shows a selection of the models that existed when this article was published. Opus 4.7 sits at 2,048 tokens, Opus 4.6 and 4.5 at 4,096; Mythos 5 sits at 512, as do Opus 5.5, Sonnet 5.5 and Haiku 5.5, which have been added since.

The difference matters in practice. Take a chatbot with a system instruction of 3,000 tokens, 40,000 requests a month, 200 tokens of user input and 300 tokens of response, counted on Haiku 4.5. There the prefix falls below the 4,096-token threshold and is therefore billed in full on every single request. Sonnet 5 counts around 30 percent more tokens for the same text but can cache the prefix; as in the first example, this assumes 1-hour retention and 252 write operations a month.

Sonnet 5 costs twice as much per token and counts more tokens, yet ends up only around 24 US dollars above Haiku 4.5. Of its price advantage, Haiku 4.5 keeps only a small remainder here, and that is down to the minimum length alone.

There is a way out, though: extend the prefix to 4,200 tokens and it crosses the threshold and becomes cacheable. The bill on Haiku 4.5 then falls to 86.92 US dollars. In this case a longer prompt costs less than a shorter one, because it falls into a different tariff.

What quietly invalidates the cache

A cache entry only holds while the beginning of the request stays exactly the same. Changes take effect hierarchically: tool definitions, then system instruction, then messages. Changing one level invalidates it and everything below.

These triggers invalidate the cache, and none of them looks like a cost decision:

• A changed tool description in the tools array. A new field in a tool schema invalidates everything that follows. On Fable 5 and 5.1, Mythos 5 and 5.1, Opus 4.8, 5 and 5.5, Sonnet 5.5 and Haiku 5.5, tools can instead be added and removed through a system message in the conversation (beta), and since 22 September 2026 also changed. The cache then stays intact.
• Turning web search or citations on or off. Both alter the prefix.
• An image added or removed. Regardless of where in the prompt it sits.
• A changed thinking or effort level. That belongs to the cached frame as well. On Fable 5.1, Mythos 5.1, Opus 5, Opus 5.5, Sonnet 5.5 and Haiku 5.5, the effort level can be changed through a system message in the conversation (beta) without losing the cache.
• A changed tool choice. Switching between automatic selection and tools turned off is enough. Fable 5.1, Mythos 5.1, Opus 5.5 and Sonnet 5.5 reject forced selection with an error anyway.

One mistake is even more mundane. If a timestamp, a customer name or a case number sits at the start of the system instruction, the prefix differs on every call. Then a hit never occurs, and the bill looks like the row without caching above. Variable details like these belong after the fixed part, not before it.

Something that can look risky is in fact allowed: appending new messages to a running conversation invalidates nothing, as long as the breakpoint sits on the immutable part. More than four explicit breakpoints per request, however, are not possible.

What else matters for model choice

Two further points belong in the picture, and the first concerns a model that chatbots and phone assistants actually run on.

Sonnet 5 stays at 2 and 10 US dollars. The price was announced as introductory pricing through 31 August 2026, after which it was to rise to 3 and 15 US dollars on 1 September. Anthropic cancelled the increase back in August (the release notes date this to 10 August 2026); the introductory price has been the standard price since. For the workflow calculated above that means 383.04 US dollars less per month, a third of the bill that the announced price would have produced.

At release, Fable 5.1 carried the most recent knowledge cutoff in the catalogue. It ends in June 2026, against May 2026 for Opus 5, January 2026 for Sonnet 5 and February 2025 for Haiku 4.5. That was one of the three questions left open in the article on the announced release; it now has an answer. Opus 5.5, Sonnet 5.5 and Haiku 5.5 now also reach June 2026.

On benchmarks, the gap to the predecessor is largest exactly where workflows run autonomously across many steps: 55.8 against 42.0 percent on Terminal-Bench 4.0 and 31.4 against 17.1 percent on AutomationBench. Part of the AutomationBench gap is likely due to the safeguards: where they intervened, Anthropic scored Fable 5 at zero. On multidisciplinary reasoning the jump is smaller, for example 60.9 against 57.8 percent on Humanity's Last Exam without tools. That matches the price change: the same field of use was both improved and made cheaper.

One detail for comparisons across generations: since Claude 4.7 Anthropic has used a new tokenizer that, per the documentation, produces roughly 30 percent more tokens for the same text. Haiku 4.5 still uses the old one. The price gap between Haiku 4.5 and Sonnet 5 is therefore wider per unit of text than the input column suggests.

What does this mean for model choice?

Four rules follow from these numbers, and they still hold when a different model appears tomorrow.

First: check the architecture before you change the model. A workflow without caching pays for the fixed prefix in full on every call; in the example, the workflow on Fable 5.1 costs 9.3 times as much without caching as with it. When this article was published, that was more than any announced discount; since Haiku 5.5 there has been a version jump of the same order (see the update).

Second: calculate with your own volumes. The shares shift with the size of the prefix and the number of steps. A workflow with a short prefix and long answers has its money in the output, and there is no cache discount there.

Third: switch to Fable 5.1 if you run on Fable 5. The list price stays the same while the cache price drops. Check three breaking API changes first: forced tool choice returns an error, only Fable 5.1 and Mythos 5.1 read Fable 5.1 thinking blocks, and editing earlier turns invalidates thinking blocks. Anyone running on Sonnet 5 or Haiku 4.5 gains nothing from this discount, because the 0.025 multiplier applies only to Fable 5.1 and Mythos 5.1.

Fourth: keep an eye on the minimum length. It can make a cheap model expensive without anyone having made a mistake. The 100,000-token threshold on Haiku 5.5 works in a similar way: above it, every line item costs five times as much (see the update).

For most mid-market applications the recommendation therefore stayed what it was before 1 September: chatbots, phone assistants and back-office workflows run well enough on Sonnet 5 or Haiku 4.5 at a fraction of the cost. The premium for Fable pays off where a workflow has to run across many steps without supervision. Both models now have successors; more on that in the update. How such a workflow is built and its benefit evidenced is covered in the article on back-office process automation.

How do you tell whether your setup is cache-friendly?

These six questions can be settled with your own team without developer knowledge. Any question that makes you hesitate may point to a cost item left open.

1. Does the immutable part of the request come first? Instruction, rules and examples first, variable details afterwards.
2. Is there a timestamp or a name in front of the fixed part? If so, a hit never occurs.
3. Does the fixed part reach the minimum length of the model in use? Between 512 and 4,096 tokens, depending on the model.
4. Does the retention period match the request rhythm? If requests arrive every minute, five minutes is enough. With gaps of half an hour the one-hour retention is cheaper despite the higher write price.
5. Are tool definitions in the tools array changed at runtime? Every such change invalidates the entire cache.
6. Does the billing show cache hits? The API response contains the fields for cache tokens read and written. If they show zeros, caching is not taking effect.

The last question is the most important, because it makes the other five unnecessary. The share of cache hits in total input is the metric an automation workflow can be measured on. When it rises, the bill falls, without anything having been changed about the model.

Update: What has changed since 2 September?

As of 14 October 2026. Since this article was published, Anthropic has released three models, and two of them have their own, lower cache price.

All figures per million tokens. On Sonnet 5.5 the cache price of 0.10 US dollars has applied since 7 October 2026; before that it was 0.20. The Haiku 5.5 prices apply to requests up to 100,000 tokens; above that, every line item costs five times as much.

Cache hits therefore come in three levels: 2.5 percent of the input price on Fable 5.1 and Mythos 5.1, five percent on Opus 5.5 and Sonnet 5.5, and ten percent on all other models, including Haiku 5.5. A lower cache price is no longer reserved for the most expensive tier.

All three models use the new tokenizer, so the volumes of the back-office workflow apply unchanged:

Where Haiku 5.5 is up to the task, this shifts the ranking of the levers in two places. On Haiku, the version jump does more in the example than caching: without caching the workflow costs 216.72 US dollars on Haiku 5.5, less than on Haiku 4.5 with caching. For requests up to 100,000 tokens, Haiku 5.5 costs 90 percent less per token than Haiku 4.5. That does not make caching redundant: on Haiku 5.5 it still brings the bill down from 216.72 to 38.30 US dollars, to less than a fifth.

In addition, the model tier overtakes the architecture: Sonnet 5.5 costs 564.48 US dollars with caching, Haiku 5.5 38.30. Moving from Sonnet 5.5 to Haiku 5.5 cuts the bill by a factor of 14.7, caching on Sonnet 5.5 only by a factor of 7.7. Between architecture and model version the order holds on Opus and Sonnet; there the version jump lowers the bill by a factor of 1.7 and 1.4.

On model choice, Anthropic now recommends starting with Opus 5.5 when in doubt; by Anthropic's account, Opus 5.5 performs at the level of Fable 5.1 on most work. Anthropic recommends Fable 5.1 for demanding reasoning and long-horizon agentic work, or when Opus 5.5 still falls short at higher effort. In its models overview, Anthropic now lists Sonnet 5 and Haiku 4.5 as legacy models. Both remain available and have not been deprecated; Anthropic gives at least 60 days' notice before a retirement. According to the documentation, Mythos 5.1 is now open to organisations that have passed one of Anthropic's verification programmes, such as the Cyber Verification Program.

Frequently asked questions

Sources

• Anthropic (1 September 2026): Introducing Claude Fable 5.1 and Claude Mythos 5.1 - release, cache price, stated savings, Claude Mythos 5.1, benchmark figures and their categories
• Anthropic (22 September 2026): Claude Opus 5.5 - performance at the level of Fable 5.1 on most work
• Anthropic (7 October 2026): Claude Haiku 5.5 - 90 percent cheaper than Haiku 4.5 for requests up to 100,000 tokens
• Claude documentation: Pricing, retrieved on 2 September and 14 October 2026 - complete price table, the 0.025 and 0.05 multipliers for cache hits, the introductory price of Sonnet 5 as the standard price, around 30 percent more tokens from the new tokenizer
• Claude documentation: Release notes, retrieved on 14 October 2026 - Sonnet 5 price standard since 10 August 2026, releases of Opus 5.5, Sonnet 5.5 and Haiku 5.5, Sonnet 5.5 cache price since 7 October 2026, beta features for tool and effort changes
• Claude documentation: Models overview, retrieved on 2 September and 14 October 2026 - knowledge cutoffs of Fable 5.1, Opus 5.5, Sonnet 5.5 and Haiku 5.5, the guidance on model choice and the listing of Sonnet 5 and Haiku 4.5 as legacy models
• Claude documentation: model pages Claude Fable 5, Claude Opus 5, Claude Sonnet 5 and Claude Haiku 4.5, retrieved on 14 October 2026 - knowledge cutoffs of the earlier models
• Claude documentation: Claude Haiku 5.5, retrieved on 14 October 2026 - new tokenizer and minimum cacheable length of 512 tokens
• Claude documentation: What's new in Claude Fable 5.1, retrieved on 14 October 2026 - three breaking changes compared with Fable 5
• Claude documentation: Prompt caching, retrieved on 2 September and 14 October 2026 - minimum cacheable lengths, retention periods, the number of permitted breakpoints and the triggers that invalidate a cache entry
• Claude documentation: Mid-conversation system messages and Effort, retrieved on 14 October 2026 - changing tools and the effort level mid-conversation without losing the cache
• Claude documentation: Model deprecations and Claude Mythos 5.1, retrieved on 14 October 2026 - status of Sonnet 5 and Haiku 4.5, notice period before a retirement, access to Mythos 5.1

The monthly bills are our own calculations based on these prices. The assumptions are stated in the tables and the sentences below them, so that they can be recalculated with your own volumes. They contain no discounts, no premiums for regional endpoints and no batch discount.

This article was published on 2 September 2026. On 14 October 2026 we revised it and corrected statements that were already wrong at publication: Anthropic cancelled the Sonnet 5 price increase back in August, not on 1 September. The Haiku 4.5 versus Sonnet 5 counter-calculation and the Haiku column of the back-office bill ignored that Haiku 4.5 splits the same text into fewer tokens. The introduction and the preview text in the blog overview ranked the cache line differently from the article's own ranking of the levers. Anthropic files Humanity's Last Exam under multidisciplinary reasoning, not knowledge tasks. The minimum lengths, the assumption on write operations, the cache invalidation triggers, the changes when moving to Fable 5.1 and the context for the AutomationBench gap were incomplete. We removed generalisations and a threshold in the checklist that had no evidence behind them. The write prices of the two retention periods are now stated correctly. The update on Opus 5.5, Sonnet 5.5 and Haiku 5.5 is new.