Four levers lower the bill, and the Claude price list of 18 October 2026 publishes a multiplier for each one: match the model to the task, cache repeated context, run work that is not time sensitive through the Batch API, and track token usage[1]. Caching and batch discounts combine, so you apply both[1].

This is written for an operations manager or COO who already has an AI bill running and wants it lower without a rebuild. Every number below comes from the published price list, with the date it was read.

How do you lower the bill of a language model?

By pulling the four levers the documentation itself names, in an order that gives you numbers before it gives you changes. Anthropic lists them as using appropriate models, implementing prompt caching, batching operations, and monitoring usage patterns[1]. Each lever carries a published multiplier, so the saving can be worked out in advance.

  • Match the model to the task. Haiku for simple tasks, Sonnet for most production workloads, Opus for the most complex reasoning[1].
  • Cache repeated context. A cache hit costs a fraction of the standard input price[1].
  • Batch what is not time sensitive. The Batch API processes large volumes of requests asynchronously at a 50% discount on both input and output tokens[1].
  • Track token usage. Monitoring token consumption is how you find out which of the first three is worth pulling[1].

What this article does not do is add up what an automation costs in the first place. That is in how much AI automation costs, and the monthly bill of a single agent is worked out in what an AI agent costs per month. Here the bill already exists and the question is which knob moves it.

What does each lever save on the 18 October 2026 price list?

The multipliers are published, so this table repeats them rather than estimating. Every row names the source it comes from, and the right-hand column is the part most summaries leave out: the case where the lever does nothing.

LeverWhat the price list publishesWorks onWhen it does not help
Match the model to the taskClaude Haiku 5.5 $0.10 and $0.50 per million tokens for prompts up to 100,000 tokens, Claude Sonnet 5.5 $2 and $10, Claude Opus 5.5 $4 and $20[1]input and outputwhen the task needs the reasoning the smaller model does not have, or when the two models run different tokenizers[1]
Prompt caching, the write1.25x the base input price for the 5-minute cache, 2x for the 1-hour cache[1]inputwhen the context changes on every call, so you pay for a write that is never read
Prompt caching, the hit0.1x the base input price, and 5% of the standard input price on Claude Opus 5.5 and Claude Sonnet 5.5, 2.5% on Claude Fable 5.1 and Claude Mythos 5.1[1]inputbelow one read on the 5-minute cache or two on the 1-hour cache, where the write is not yet earned back[1]
Batch API50% discount on both input and output tokens[1]input and outputwork that has to answer now, and sessions, which have no batch mode[1]
Track token usageno multiplier; named as the fourth lever, to identify optimization opportunities[1], counted per model ID with token counting[1]measurement, not pricewhen nothing is logged, there is nothing to find
Session runtime on agents$0.08 per session-hour, counted only while the session has the status running[1]time, alongside tokenstime spent idle, rescheduling or terminated, which does not count toward runtime[1]

How much does prompt caching actually change?

Writing to the cache costs 1.25x the base input price for the 5-minute duration and 2x for the 1-hour duration, and a cache hit costs 0.1x the base input price[1]. On Claude Opus 5.5 and Claude Sonnet 5.5 a hit costs 5% of the standard input price instead of 10%, and on Claude Fable 5.1 and Claude Mythos 5.1 it costs 2.5%[1].

That makes caching an arithmetic question, not a preference. The documentation states the break-even itself: caching pays off after one cache read for the 5-minute duration and after two for the 1-hour duration[1]. So the thing to check is not whether your context is large, but whether the same context gets read back inside the window. A system prompt, a product catalogue or a policy document pulled in through an MCP server on your own systems usually does. A customer email does not.

Do the batch discount and caching stack?

They do, and the documentation says so twice. The cache multipliers “stack with other pricing modifiers, including the Batch API discount and data residency”, and batch and prompt caching discounts can be combined[1]. Most pages on this subject present the two as alternatives, which quietly costs you the smaller of the two.

One exception matters. Claude Managed Agents sessions are stateful and interactive and there is no batch mode for them, so the Batch API discount does not apply to a session[1]. Batch stays a lever for queued, non-interactive work: nightly classification, bulk enrichment, backfills.

Why is a cheaper model not always cheaper?

Because the two models may not count your text the same way. Claude 4.7 and later models, and Claude Mythos Preview, use a newer tokenizer that produces approximately 30% more tokens for the same text; the exact increase depends on the content and the workload shape, and Claude Sonnet 4.6 and earlier models use the previous tokenizer[1].

A price per million tokens is therefore only half of a comparison. The other half is how many tokens your own text becomes on that model, which depends on the tokenizer, the language and the type of content[1]. The documentation gives one way to settle it instead of guessing: count the request with token counting and that model’s ID[1]. Run that on a real prompt from your workload before you move anything, because the published 30% is an approximation and your own swing can sit either side of it.

What does the worked example in the documentation show?

The price list works out one case, and this is that case as it stands. A one-hour coding session on Claude Opus 5.5 consuming 50,000 input tokens and 15,000 output tokens: input tokens 50,000 x $4 / 1,000,000 = $0.20, output tokens 15,000 x $20 / 1,000,000 = $0.30, session runtime 1.0 hour x $0.08 = $0.08, total $0.58. With prompt caching active and 40,000 of those input tokens served as cache reads: uncached input tokens 10,000 x $4 / 1,000,000 = $0.04, cache reads 40,000 x $4 x 0.05 / 1,000,000 = $0.008, output tokens $0.30, session runtime $0.08, total $0.428[1].

Two things are worth reading off it. The runtime line does not move, because Claude Managed Agents is billed on two dimensions, tokens and session runtime, at $0.08 per session-hour, measured to the millisecond and counted only while the session has the status running; idle time, rescheduling and terminated sessions do not count[1]. And the output tokens do not move either, because caching is an input lever. Getting the expensive context cached and the long waits out of a running session is build work on the agent itself, which is what WeAdapt does as custom AI tools.

In which order do you pull the levers?

Measurement first, model choice second, and the two discounts last. That order is deliberate: the discounts are the easy part, but without usage data you cannot tell which workload is carrying the bill, and which workloads are worth reworking is a consultancy question before it is an engineering one.

  1. Log token usage per workflow before changing anything. Monitoring usage patterns is the lever that tells you where the other three apply[1].
  2. Count tokens on the model ID you actually run, with token counting rather than a rule of thumb[1].
  3. Check the tokenizer generation of every model in the comparison, since 4.7 and later count roughly 30% more tokens for the same text than Sonnet 4.6 and earlier[1].
  4. Match each step to a model. Haiku for the simple steps, Sonnet for most production workloads, Opus where the reasoning has to be there[1].
  5. Cache the context that repeats, and pick the 5-minute or 1-hour duration on how soon the next read lands[1].
  6. Move queued work to the Batch API for the 50% discount on input and output, and combine it with caching where both apply[1].
  7. Track session runtime separately on agents, because runtime accrues only in the running status and the Batch API discount does not reach it[1].

A lower bill only means something next to what the workflow brings in. The two formulas for that side of it are in calculate the ROI of an AI automation.

Want the bill of an AI workflow that is already running to come down without rebuilding it? WeAdapt looks at the model per step, the caching windows and the batch queues inside the systems you already have. Book a call, or read the FAQ first.

Sources

  1. Anthropic, Pricing (platform.claude.com, accessed 18 October 2026)