Guides / anthropic api cost
Anthropic API cost and the cache-write charge
Anthropic API cost runs from $1.00 per million input tokens on Haiku to $10.00 on Fable, and every single model bills output at exactly five times input. Anthropic is also the only provider here charging to write a cache. Culpa, a local-first LLM cost, margin, and forecast ledger, meters cached, uncached and cache-write tokens separately per call.
Why this happens
Two facts define Anthropic's economics. The output-to-input ratio is exactly five on all ten models, so the shape never changes as you move tiers and only the absolute rate does. And cache writes bill above the input rate, at 1.25 times for the five-minute window and 2 times for the one-hour window. That second fact makes caching a decision with a downside, which is unusual, and most teams switch it on without checking the ratio.
What this usually looks like
- Spend rose slightly after enabling prompt caching, which is the signature of writes outnumbering reads.
- Cache hits collapse after every deploy, because the cached block is edited as part of routine tuning.
- A tier upgrade cost exactly what you expected on input and five times that on output, which surprised somebody.
- Your cached-token share is near zero despite an identical system prompt on every call.
Free, no card, no account
Run the free Cost Leak Scan
It shows your most expensive conversation before you install anything.
Mistakes that cost the most
| Mistake | Why it hurts | Do instead |
|---|---|---|
| Enabling caching without checking the write-to-read ratio. | A cache write bills at 1.25 times input, so churning it costs more than never caching at all. | Count writes against reads. A five-minute cache clears break-even on the first read, so switch caching off only where writes never get read. |
| Budgeting a tier change on the input rate alone. | Output is exactly five times input on every model, so the output side moves five times as hard. | Model both sides at the new rate using your own token split before approving the change. |
| Editing the cached system prompt as part of routine tuning. | Each edit invalidates the prefix for all traffic at once and forces a paid rebuild. | Freeze the cached block and put anything you tune often after it, where changes cost nothing. |
Run this check tonight
- Divide cached input tokens by total input tokens to get your hit rate.
- Count cache writes against reads over the same window. That ratio decides whether caching pays here.
- Check your output-to-input token split, then multiply the output side by five to see where the bill sits.
- Plot hit rate against deploy timestamps. Cliffs at deploys mean the cached block is being edited.
- List which Anthropic tier each feature uses. All ten models share one price shape, so the tier is the decision.
When the cache write turns a saving into a loss
A 5,000-token system prompt on Claude Sonnet 5 at its introductory rates, $2.00 per million input and $0.20 cached read. The five-minute cache write bills at $2.50 per million, derived at 1.25 times input.
Caching wins from the first read. A one-hour write at 2x needs two reads. Only a write nobody ever reads loses, and then by 25%.
Every number, with its confidence and source
| Figure | What it means | Confidence | Source |
|---|---|---|---|
| 5.0x | output-to-input ratio, identical on all ten Anthropic models | calculated | Output divided by input on every anthropic row in the price book, effective 2026-07-02. All ten return exactly 5. |
| 1 read | break-even point before a five-minute cache write pays for itself on Sonnet 5 | calculated | Incremental write cost $0.0025 divided by the $0.009 per-read saving = 0.28, so the first read clears it. Real Sonnet 5 introductory rates. Matches Anthropic's published statement that a five-minute cache pays off after one read and a one-hour cache after two. |
What a generic answer can’t know
A pricing page lists a cache-write rate. It can't tell you how often your cache is rebuilt versus read, because that depends on your traffic pattern and how often you deploy. Culpa measures both on your infrastructure, keeps your prompts and responses there, and counts the calls to run your plan.
Questions founders ask next
How much does Anthropic charge to write a cache?
1.25 times the input rate for the five-minute window and 2 times for the one-hour window. On Sonnet 5 at $2.00 per million input that's $2.50 and $4.00. Both sit above the input rate, so a churning cache costs money.
Do all Anthropic models have the same output ratio?
Yes, exactly five times input across all ten models in this price book, from Haiku at $1.00 and $5.00 to Fable at $10.00 and $50.00. The shape never changes, only the absolute rate does.
Which Anthropic model is cheapest?
Claude Haiku 4.5 at $1.00 per million input and $5.00 output, with cached reads at $0.10. It sits ten times below Fable on input, and the whole line shares the same five-to-one output shape.
On your infrastructure
Culpa runs on your infrastructure. Your prompts and responses never leave it. Culpa counts calls to run your plan, and it fails open, so if it ever breaks your app keeps running.
Free, no card, no account
Run the free Cost Leak Scan
It shows your most expensive conversation before you install anything.
Keep reading
Sources: Anthropic pricing. Last reviewed 2026-08-01, rates effective 2026-07-02. Plain text version.