Your Anthropic bill usually isn't high because Claude is expensive. It's high because output tokens, long contexts, and un-cached prompts compound quietly — and you find out a month later on the invoice.
This is for engineering leaders and founding engineers whose last Anthropic invoice came in higher than expected. The goal is simple: name the likely cause, run a structured check, and decide whether to fix now or add monitoring so the next spike shows up on day one instead of at month end. (If your Claude spend is a subscription seat cost rather than API usage — Claude Code, for example — see Claude Code cost options and overage instead; this article is about the metered API bill.)
What usually drives a high Claude bill
Most high Anthropic bills come from one of four things, and usually a combination:
- Output-heavy workloads. Claude bills output tokens several times higher than input. A workload that generates long completions — drafting, code generation, agents that "think out loud" — pays far more per call than the token count alone suggests. See input vs output tokens for why the split matters.
- Long context, uncached. Sending a large system prompt or retrieved documents on every call is expensive when it isn't cached. Prompt caching bills reused context at a fraction of fresh input — leaving caching off is one of the most common avoidable costs.
- A model-tier jump. Moving traffic from a mid-tier model to a flagship for quality can multiply per-token cost — the tier gaps are visible in the LLM API Pricing Index. Often only a slice of that traffic actually needed the bigger model.
- Volume you didn't see. Retries, background jobs, and agent loops quietly multiply call counts. A single retrying agent can double a workload's tokens without any change to user-facing behaviour.
How to investigate
Work from the shape of the spend, not a guess:
- Split by direction first. Look at input vs output vs cache tokens per model. If output dominates, the fix is prompt and output-length discipline, not a model swap. If input dominates and cache is near zero, caching is your lever.
- Attribute by model and workload. One model or one feature usually drives the jump. Group usage by model and project so you're optimising the 80%, not the noise.
- Check your blended rate. Your real blended token rate — input, output, and cache combined at your actual mix — is the only fair basis for comparing models or forecasting a change.
The AI Explorer does exactly this across every provider: token usage by model, project, and user, with the input/output/cache drill-down behind each line of the bill. If you're mid-incident, the how to investigate an AI spend spike runbook is the fast path.
How to cut it
- Turn on prompt caching for any stable system prompt or shared context. This is often the single largest, lowest-risk saving.
- Right-size the model per workload. Move the traffic that doesn't need a flagship to a cheaper model that still clears your quality bar — see how to switch to cheaper AI models without losing quality. Model Recommendations automates this: it surfaces cheaper models that match or beat your current one on published benchmarks, priced at your real token mix, so you don't downgrade quality to save money.
- Cap output and trim context. Set
max_tokensdeliberately, and stop sending context the model doesn't use. - Prevent the next spike. Anthropic cost monitoring with anomaly detection gives you daily visibility and an alert the day spend moves — not four weeks later.
FAQ
Why is my Claude API bill so high?
Usually because output tokens (billed several times higher than input), long uncached contexts, a model-tier jump, or hidden volume from retries and agent loops are compounding. Split your usage by input/output/cache tokens per model to see which one is driving it before you change anything.
How do I reduce Claude API costs without losing quality?
Turn on prompt caching for stable context, cap output length, and move only the traffic that doesn't need a flagship model to a cheaper one that still clears your quality bar on published benchmarks. The LLM cost calculator shows what a candidate switch is worth at your token mix. Validate any switch with a guarded trial rather than moving all production traffic at once.
How do I see Claude token usage by model?
Connect Anthropic to StackSpend and the AI Explorer shows token usage by model, project, and user, with the input/output/cache split — so you can see which workload and which token direction is driving the bill.
Related
- AI Explorer — token usage by model, with input/output/cache behind every line
- Model Recommendations — cheaper models that clear your quality bar, priced at your mix
- Anthropic Cost Monitoring — track Anthropic spend by model and project
- Why Your OpenAI Bill Is So High — the same investigation for OpenAI
