237€ of Claude API in March. Here's where the money went.

I paid 237.40 € for Claude API in March 2026. Not Claude Code (my Max subscription is separate), not Claude.ai. Just the raw API, which I use in homemade agents, batch scripts, and an internal tool. I wanted to lay this out because I see a lot of devs throwing around “AI is cheap” without ever showing actual numbers.

Here’s my month, honestly, line by line.

The real breakdown

I exported the invoice from the Anthropic dashboard. Three line items dominate:

Item	Cost	% of total
Long-running agents (cron, overnight tasks)	142 €	60%
Batch scripts (doc import, embeddings, classification)	54 €	23%
Interactive requests (personal CLI, REPL)	41 €	17%

60% of my bill comes from three agents running continuously. One monitors and triages my GitHub issues, one compiles daily technical newsletters, and one prepares log summaries for my studio. All three run on Sonnet 4.6, with prompt caching.

What prompt caching saved

Without prompt caching, the same bill would probably sit around 390 €. The math is easy: my agents systematically load ~25K tokens of context (codebase, instructions, condensed history) on every turn. Without caching, every call pays full price for those 25K tokens. With caching, I save ~90% on that portion.

Concretely: running a Sonnet 4.6 call on a 25K-token cached prompt plus 1K tokens of delta costs roughly 0.008 €. Without caching, it’s 0.075 €. That difference gets brutal when you’re hitting 200 calls a day.

Lesson: if you’re building an agent that repeatedly calls back to the same context base, caching isn’t optional — it’s the difference between economically viable and not.

What cost a lot and wasn’t worth it

Three usages where I burned money for zero ROI:

1. Document classification on Opus 4.7 — 38 € for nothing. I was sorting internal PDFs. I started on Opus 4.7 “out of habit.” After benchmarking it, Haiku 4.5 did the exact same task at 4% of the price with an equivalent error rate (0.8% vs 0.6%, not statistically significant over 500 docs).

2. A “smart” agent that summarized emails 15 € to summarize 2000 emails. Usage: zero. Nobody read the summaries. I killed the agent, kept a simple regex filter plus manual extraction instead.

3. Some experiments on Opus 4.7 in agentic tool use About 30 €. Not wasted, that’s learning. But worth adding to the mental budget when you’re prototyping.

What’s worth every cent

The overnight cron agents — 142 € well spent. I measured it: those three agents save me ~6h/week, so ~24h/month. At my day rate, 142 € is the equivalent of 20 minutes of billing. Obvious ROI.

The log analysis batches When one of my studio’s sites has an incident, I can throw 10,000 lines of logs at Sonnet 4.6 and get a diagnosis back in 30 seconds. It saved me a weekend at Christmas. That alone is worth the 54 €/month.

The Anthropic vs OpenAI gap

Out of curiosity, I redid my main tasks on GPT-5.3 Codex for a week to compare. Three observations:

1. OpenAI’s pricing is more aggressive on the equivalent model. Roughly 15-20% cheaper at equal token counts.
2. OpenAI’s caching is less flexible: less granularity on what gets cached, so in practice fewer savings.
3. Agentic quality still lags behind Claude on my use cases: following long instructions, multi-step work, nested tool use. For one-shot tasks, it’s close.

Conclusion: I’m sticking with Anthropic for anything agentic. I switch to OpenAI when doing one-shot batch work where price is the only thing that matters.

What this teaches me for April

My plan for this month:

• Migrate 80% of my volume to Haiku 4.5 wherever quality is equivalent. Goal: bring the bill down to 120-150 €.
• Kill the “email summary” agent — useless.
• Better cost logging per agent with a custom header, to get per-use-case ROI next month.

I’ll do a post-mortem at the end of the month to see if the optimization holds.

Bottom line

If you’re getting into Claude API agents in 2026:

• Budget 200-300 €/month as soon as you have more than two agents in production doing more than a demo.
• Caching is not optional. Without it, your ROI drops by a factor of 3.
• Don’t launch anything on Opus 4.7 without testing Sonnet and Haiku first. The “bigger = better” reflex gets expensive.
• Measure ROI per agent, not globally. Some agents in my bill cost 5x their value, others 20x.

AI is a variable cost. It’s not “the 20 €/month subscription.” It’s a line item that can explode if you don’t watch it. And that’s fine — it’s just cloud bill management, like AWS ten years ago.