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 appears in none of 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 weighs more than the jump between two model generations. 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 will not happen.

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 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. Across the entire Claude price list, a cache hit costs ten percent of the input price. Fable 5.1 and Mythos 5.1 are 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.

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 is it in no comparison?

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 at 1.25 times the write price and one hour at twice.

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. The cache price sits one click deeper.

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.

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 received no special cache price. The discount narrows the gap between the tiers, it does not reverse it.

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, and for most companies it runs the surprising way round:

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 and the one that gets written about most.

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.

Why the cheapest model can produce the highest bill

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 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. On Haiku 4.5 the prefix falls below the 4,096-token threshold and is therefore billed in full on every single request.

The model with half the list price costs 24.98 US dollars more per month, or 15.3 percent. Comparing only the input column leads to the wrong choice here.

There is a second, even less obvious way out: 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. A new field in a tool schema invalidates everything that follows.
• 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.
• A changed tool choice. Switching between automatic and forced selection is enough.

The most common mistake is 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.

What many consider 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 changed on the same day

Two further changes matter more to mid-sized applications than the Fable discount, because they concern the models 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. That increase will not happen; the introductory price has become the standard price. For the workflow calculated above that means 383.04 US dollars less per month, a third of the bill. This news made no headline.

Fable 5.1 carries 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.

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. On knowledge tasks 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 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 is made cheap by no model. The factor of 9 is larger than any discount announced so far.

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. List price and behaviour stay largely the same while the cache price drops. Anyone running on Sonnet or Haiku gains nothing from this discount, because it applies exclusively to Fable 5.1 and Mythos 5.1.

Fourth: keep an eye on the minimum length. It is the only value in this article that can make a cheap model expensive without anyone having made a mistake.

For most mid-market applications the recommendation therefore stays 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. 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. If more than two of them cause hesitation, there is a cost item lying 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? 512, 1,024 or 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 doubled write price.
5. Are tool definitions changed at runtime? Every change invalidates the entire cache below it.
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.

Frequently asked questions

What is a cache hit and why does it cost money? A cache hit occurs when a model processes a piece of text it has already processed in an earlier request and that has stayed unchanged since. The provider does not have to compute that part again and passes the advantage on in the price. It is not free: on Claude models a cache hit costs ten percent of the normal input price, and on Fable 5.1 and Mythos 5.1 only 2.5 percent. The cached text also has to be written once, and that costs more than normal input.

How much does the new Fable 5.1 cache price actually save? Anthropic states around 25 percent for typical workloads and up to about 45 percent for highly agentic ones. In our worked monthly bill for 4,200 cases with eight model calls each and a fixed prefix of 60,000 tokens, the bill falls from 3,830.40 to 2,318.40 US dollars, a reduction of 39.5 percent. The figure sits at the upper end of the range because the prefix in this setup is very large.

Does the cache discount make Fable 5.1 the cheapest model now? No. In the same calculation Opus 5 costs only 1,915.20 US dollars and Sonnet 5 just 766.08, because both are considerably cheaper on input and output. The cache discount narrows the gap between the tiers, it does not reverse it. Fable remains the tier for workflows that have to run across many steps without supervision.

Why can a cheaper model produce the higher bill? Because the minimum cacheable length differs by model. It is 512 tokens on Fable 5.1, Mythos 5.1, Fable 5 and Opus 5, 1,024 on Sonnet 5 and 4,096 on Haiku 4.5. A system instruction of 3,000 tokens cannot be cached on Haiku at all. In our counter-calculation the same chatbot therefore costs 188.00 US dollars a month on Haiku 4.5 and only 163.02 on Sonnet 5, even though Haiku carries half the list price.

What invalidates the cache without anyone noticing? Changes take effect hierarchically, from tool definitions through the system instruction to the messages. Adjusting a tool description invalidates everything that follows. The same applies to turning web search or citations on or off, to images added or removed anywhere in the prompt, to changed thinking or effort levels and to a changed tool choice. A timestamp or a customer name at the start of the system instruction is enough to ensure a hit never occurs.

Does Sonnet 5 stay at 2 and 10 US dollars? Yes. The 2 US dollars per million input tokens and 10 US dollars per million output tokens were originally announced as introductory pricing through 31 August 2026, and have been the standard price since 1 September. The announced increase to 3 and 15 US dollars will not happen. For the workflow calculated in this article that is 383.04 US dollars less per month.

What changed about the knowledge cutoff of Fable 5.1? The reliable knowledge cutoff ends in June 2026, where Fable 5 ended in January 2026 and Opus 5 in May 2026. Fable 5.1 therefore carries the most recent knowledge cutoff in the Claude catalogue. For workflows that supply their own documents this matters less than often assumed, because there the supplied sources are what counts.

Is it worth switching to Fable 5.1 straight away? Only if you already run on Fable 5. The switch is then cost-neutral to cheaper, because list price and behaviour stay largely the same while the cache price drops. Anyone running on Sonnet or Haiku today gains nothing from the discount: the lower cache price applies exclusively to Fable 5.1 and Mythos 5.1.

Sources and status

The details of the release, the cache price, the stated savings and Claude Mythos 5.1 come from Anthropic's announcement of Claude Fable 5.1 and Claude Mythos 5.1 dated 1 September 2026. The benchmark figures come from the same source.

The complete price table, the 0.025 multiplier for cache hits on Fable 5.1 and Mythos 5.1, and the note that the introductory price of Sonnet 5 has become the standard price come from the Claude platform pricing page, retrieved on 2 September 2026.

Context windows, knowledge cutoffs, model IDs and retirement dates come from the models overview in the same documentation, also retrieved on 2 September 2026. That page also carries the note on the tokenizer in use since Claude 4.7.

Minimum cacheable lengths, retention periods, the number of permitted breakpoints and the list of triggers that invalidate a cache entry come from the prompt caching documentation page, retrieved on 2 September 2026.

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

How the release of Fable 5.1 was evidenced in advance and which questions remained open at the time is covered in the article from 31 August. How a model change affects running applications is shown in the article resolving the Ox Alpha case using a completed example.