Short answer: Anthropic lists Claude Sonnet 5 at $2.00 per million input tokens and $10.00 per million output tokens, with prompt-cache writes at $2.50 per million (5-minute) or $4.00 (1-hour) and cache reads at $0.20 per million. Batch processing halves those rates to $1.00/$5.00 with cache read at $0.10 and write at $1.25. Price the cache lifecycle, not just the two headline rates.
| Meter | Standard | Batch | Note |
|---|---|---|---|
| InputSonnet 5 | $2.00 | $1.00 | Batch halves input for asynchronous work. |
| OutputSonnet 5 | $10.00 | $5.00 | Output is 5x input; answer length drives the bill. |
| Cache write (5m)Sonnet 5 | $2.50 | $1.25 | A 25% premium over writing the same tokens uncached. |
| Cache write (1h)Sonnet 5 | $4.00 | $2.00 | Longer retention costs 2x the input rate. |
| Cache readSonnet 5 | $0.20 | $0.10 | 90% below input when the prefix actually hits. |
| Haiku 5.5Prompts ≤100k | $0.10 | $0.50 | A cheaper route for classification and extraction. |
Input, output, and the cache lifecycle
Anthropic's published table lists four numbers per model, and only two of them are the familiar input and output rates. Sonnet 5 is $2.00 input and $10.00 output per million tokens. Prompt caching adds a write rate of $2.50 per million for a five-minute lifetime and $4.00 for an hour, plus a read rate of $0.20. Those extra meters exist because caching is a two-step transaction: you pay to write the prefix, then you pay a discounted rate every time it is read.
A cache write costs $0.50 per million more than sending the same tokens uncached. A cache read saves $1.80 per million. Break-even therefore arrives quickly: one reuse of a one-million-token prefix costs $2.50 + $0.20 = $2.70 against $4.00 for two uncached passes, a saving of $1.30. If the prefix is reused ten times the saving is large; if it is evicted before the second request it is a net loss of $0.50 per million.
- Cache lifetime is a product decision: five minutes suits a burst of requests, an hour suits an agent session with pauses.
- A prefix that changes per user or per document will not hit; stable system prompts and tool schemas will.
- Cache reads are only visible if you log the cache-read usage field, not just total input tokens.
The economics of a long-running agent
Agent workloads accumulate input: system prompt, tool definitions, retrieved documents, and conversation history all ride in the prompt. At $2.00 per million input tokens, a 40,000-token context costs $0.08 per request before any output. Run that request 100,000 times a month and the input bill alone is $8,000. The same traffic with a 60% cache-hit rate and a stable prefix drops the cached portion to $0.20 per million, removing thousands of dollars without changing the model.
Output stays expensive regardless. At $10.00 per million tokens, a 600-token answer costs $0.006, so 100,000 answers cost $600. Ten answers per request would cost $6,000. This is why output caps and response-length instructions are the highest-leverage cost controls on a Claude deployment — more effective than switching models for many products.
Batch is a scheduling decision
The batch column halves every rate: $1.00 input, $5.00 output, $0.10 cache read, $1.25 cache write. The saving is genuine, but it only applies to work that can wait for asynchronous completion. Evaluation runs, document enrichment, moderation backlogs, and overnight summarization fit. A chat interface, an interactive editor, or a checkout flow does not.
The practical pattern is to split the architecture: interactive requests at standard rates, and a queue for everything that tolerates delay. Measure the share of traffic that genuinely belongs in the queue. Teams routinely assume 50% batch eligibility and discover, after instrumenting latency requirements, that only 15% of requests can wait. Budget from the measured share.
Worked estimate: 200,000 agent requests
Take 200,000 requests per month with 3,000 input tokens and 600 output tokens, and a measured 60% cache-hit rate on a stable prefix. Uncached input is 240M tokens at $2.00 ($480.00), cached input is 360M at $0.20 ($72.00), and output is 120M at $10.00 ($1,200.00). The monthly total is $1,752.00, or $8.76 per thousand requests, and output is 68% of it.
Move half of that traffic to batch and the blended total falls to about $1,314.00. Trim average output from 600 to 400 tokens instead, and output drops to $800.00 for a total of $1,352.00. Doing both lands near $1,013.00. The comparison shows why output length and batch eligibility belong in the same budget review as the model choice.
Choosing between Claude routes
Haiku 5.5 at $0.10 input and $0.50 output is an order of magnitude cheaper than Sonnet 5 and is positioned for classification, extraction, and routing. Using it as a triage layer in front of Sonnet 5 is often cheaper than sending every request to the larger model, provided the triage task itself passes your quality bar. Measure both the pass rate and the escalation rate; a triage step that escalates 90% of traffic adds cost without removing any.
Sonnet 5.5, listed at the same $2.00/$10.00 headline rates with a lower $0.10 cache-read rate, is a reminder that a version bump can change the cache economics without changing the sticker price. Recheck the meter table for the exact model version in your configuration, and pin that version in logs so a later price review has a reference point.
Estimate a Claude agent bill
Enter the rates for the exact Claude model and tier you use. Cache-hit share is a measured property of your prompt structure, not a default discount.
Planning result
Review the example inputs, replace each rate with the exact current value from the linked source, then calculate. APIDir does not send these values to a provider.
Questions developers ask
How much is Claude Sonnet 5 per million tokens?
Anthropic listed $2.00 per million input tokens and $10.00 per million output tokens on November 13, 2026, with cache writes at $2.50 (5-minute) or $4.00 (1-hour) and cache reads at $0.20.
Is prompt caching always worth it?
Only when the cached prefix is reused. A cache write costs $0.50 per million more than uncached input; a cache read saves $1.80 per million, so one reuse covers the premium and repeat reuses compound the saving.
Does batch processing halve Claude pricing?
Yes for eligible asynchronous work: Anthropic lists batch at $1.00 input and $5.00 output per million tokens, with corresponding cache rates. Interactive traffic cannot use those rates.
Why is my Claude bill higher than the token estimate?
Common causes are uncounted cache writes, output longer than the planned cap, retries, and tool or web-search calls billed separately at $0.01 per search on the router route.
Sources and check dates
Every rate in this guide was retrieved from the linked page on Nov 13, 2026. Prices can change; recheck the source before purchasing.