Invest1 publisher3 min readPublished
Anthropic's worked example turns a 120,000-token conversation into 2.8 million billed input tokens
Anthropic's cost breakdown for Opus 5.5 walks through a 40-turn coding task in which nine tenths of the input is resent conversation billed as cache reads. Cutting that line 60% takes the task's input bill down 39% on a 20% per-token cut.
The Investor · Invest desk
What happened
- Its worked example runs a task whose context grows from 20,000 to 120,000 tokens over 40 turns and processes about 2.8 million input tokens, because every turn resends the conversation so far.
- Those same 2.8 million tokens cost $11.20 with no caching, about $1.62 at a 90% cache hit rate and about 99 cents at 96%.
- The 60,000 output tokens of a typical task cost $1.20 on Opus 5.5, the same as six million cache reads, since one output token is priced at 100 times a read.
- Anthropic warns that lower effort, a smaller model or less context all save tokens, and that a retry costs more than those savings.
Compiled by The InvestorSomething wrong?How this is made
Why it matters
- constraint A budget written in dollars per million tokens cannot bound a Claude Code invoice, because caching alone moves one task's input cost across a range of 6.9 to one while the list price never changes.
- decision Picking an effort setting becomes a forecast of first-pass success on a particular codebase: the upside is capped by the output line, and the downside is paying for the whole task twice.
- cost Subscribers collect the Opus 5.5 cut as limits that go about 25% further, so the extra 60% off cache reads reaches only API buyers as cash.
- precedent Once a vendor publishes turn-level cost, buyers can ask every rival for cost per completed task, and a lower price per million stops settling the comparison on its own.
Back the cuts out of the new list prices and Opus 5 comes to $5 per million input tokens, $25 per million output and $0.50 per million cache reads [16]. Price the same illustrative task on those numbers, 2.8 million input tokens with nine tenths served from cache, and the input costs about $2.66 [17]. On Opus 5.5 it is $1.62 [4]. The task's input bill fell 39% on a per-token cut of 20% [18], because the line carrying most of the volume is the one Anthropic cut 60% [2].
Then the two levers a developer actually controls. Cutting the illustrative task from 40 turns to 25 saves 60 cents of input [19]. Lifting the cache hit rate from 90% to 96% saves 63 cents [19]. Of the hit rate Anthropic wrote, "No other setting moves input cost this much" [6].
Output is where the effort setting lives. Thinking is billed as output, and the customer pays for all of it even when Claude Code shows only a summary [8]. So the entire output line on this illustration is $1.20 against about $2.82 for the whole task [20]. Halve the thinking and you save 60 cents; a rerun costs the full $2.82, so the saving is gone at a retry rate of about 21%, one task in five [22]. Anthropic put it more flatly: "A retry costs more than those savings" [10].
The argument comes from the seller of the model in question, and the post says so about its own figures: "These are best effort illustrations, so be sure to check our docs and your own math" [11]. The examples bill cached input at the read price, leave out cache writes, and use illustrative token counts [12]. The post does not publish a retry rate for any effort setting [23].
I think the unit is right and this illustration is not yet evidence for it. Cost per completed task is the number a monthly budget meets, and Anthropic's suggested fix for turn count, giving the model a test or a build so it can check its own work [24], is a claim about a codebase, not about a price list. The counter-case is that on a task a smaller model or a lower effort setting finishes first time, the output saving is real and no retry ever arrives to cancel it, and the post prices the retry without pricing how often it happens [10]. What would settle it is a measured distribution of turn counts and retry rates by effort setting across real sessions.
The shape of the discount says where Anthropic expects the volume: cheaper cache reads mostly help long sessions, cheaper output mostly helps tasks that need a lot of reasoning [15], and the deepest cut went to cache reads [2]. A team running a thousand tasks a month on these figures spends about $2,820, and the difference between 40 turns and 25 is $600 of it [25].
What to watch
- Published turn counts and retry rates by effort setting, measured across real Claude Code sessions rather than built from list prices.
- Whether rival vendors answer with cost per completed task instead of price per million tokens.
- Whether the API-only cache-read cut later reaches Pro, Max and Team limits.