Guides / is my llm spend normal
Whether your AI bill is normal, and why nobody can tell you
There's no industry benchmark for LLM spend, because cost tracks what the work is rather than what the company is. The two published per-unit figures differ by 191 times. Culpa, a local-first LLM cost, margin, and forecast ledger, gives you the comparison that does work, which is your own trend and your own margin.
Why this happens
This is the first question anyone asks about an AI bill and the one with the least satisfying answer. There's no industry average worth having, and the reason is structural rather than a gap in the data. Model cost is a function of tokens, tokens are a function of what you're asking the model to do, and two companies of identical size doing different work will differ by orders of magnitude with both of them entirely healthy. The best available evidence makes that concrete. Anthropic publishes two worked examples on its own pricing page, and they sit 191 times apart per unit: an hour of agentic coding against a single support conversation. Both are correct. Neither tells you whether your number is fine. What replaces the benchmark isn't more data about other companies, it's three comparisons against yourself: your cost per unit sold over time, your gross margin with model spend inside cost of goods sold, and your cost per call against what it was last quarter. Those move for reasons you can act on. An industry average moves for reasons that have nothing to do with you.
What this usually looks like
- Somebody asked whether the AI bill is reasonable and nobody could answer.
- You've searched for an industry benchmark and found only vendor case studies.
- The bill is compared month to month and never to anything the business sold.
- Cost per unit sold isn't reported because nobody has defined the unit.
- A number was called high or low with nothing behind the judgement.
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 |
|---|---|---|
| Looking for an industry average. | Cost tracks the work rather than the company, so an average across different work means nothing. | Compare against your own trend and your own margin, which move for reasons you control. |
| Treating a vendor's worked example as a benchmark. | It prices that vendor's chosen scenario, and the two Anthropic publishes are 191x apart. | Use them to sanity-check your arithmetic, not to judge your total. |
| Comparing total spend month to month. | It rises with growth and with waste equally, so the direction tells you nothing on its own. | Divide by something: units sold, active customers, conversations resolved. |
| Judging the bill without the revenue beside it. | A large bill on a large margin is healthier than a small one on a thin margin. | Report gross margin with model spend in cost of goods sold, and track that. |
Run this check tonight
- Pick the unit your business actually sells and work out cost per unit for last month.
- Compare it to the same figure three months ago rather than to any external number.
- State your gross margin with model spend inside cost of goods sold.
- Check whether your cost per unit is rising while your price stays flat.
The only two published benchmarks, side by side
Anthropic's pricing page, read 2026-08-05, publishes two worked examples. One prices an hour of agentic coding on Claude Opus 5. The other prices support conversations on Claude Haiku 4.5. Both totals below are Anthropic's own, and re-deriving each from the rates reproduces their stated figures exactly. Every figure here is published rather than assumed.
Both numbers are right and they answer different questions. That's the whole difficulty with benchmarking this: the spread between two legitimate uses of the same vendor is wider than any plausible spread between two companies doing the same thing. An average across that range would describe nobody.
Every number, with its confidence and source
| Figure | What it means | Confidence | Source |
|---|---|---|---|
| $0.705 and $0.0037 | Anthropic's two published worked examples, an hour of agentic coding and a support conversation | provider-reported | Anthropic's pricing page at platform.claude.com/docs/en/docs/about-claude/pricing, read 2026-08-05, publishes both. The coding session is 50,000 input and 15,000 output on Claude Opus 5 plus one session-hour: $0.25 + $0.375 + $0.08 = $0.705, and $0.525 with caching. Support is about 3,700 tokens a conversation on Claude Haiku 4.5 at about $37.00 per 10,000, which is $0.0037 each. Re-deriving both from the published rates reproduces Anthropic's stated totals exactly. The ratio 0.705 / 0.0037 = 191. |
What a generic answer can’t know
Nobody outside your company knows what you sell, so nobody can tell you what a unit of it should cost. That's why the useful comparison is always internal, and why it needs cost and revenue in the same place. Culpa prices every call in exact decimal from a versioned price book, attributes it to the customer, feature and product that caused it, and takes your revenue alongside, so cost per unit sold and gross margin by segment are queries rather than quarterly assembly work. Forecasts are stored as ranges and scored against what happened, which turns your own history into the benchmark you were looking for. It's the only one that was ever going to answer the question.
Questions founders ask next
What is a normal monthly LLM spend?
There isn't one. Cost follows the work rather than the company, and Anthropic's own two published examples sit 191 times apart per unit: $0.705 for an hour of agentic coding against $0.0037 for a support conversation. An average across that range would describe no real business.
What should I compare my AI bill against?
Yourself, three ways. Cost per unit sold over time, gross margin with model spend inside cost of goods sold, and cost per call against last quarter. All three move for reasons you can act on, which no external average does.
Are vendor case studies useful at all?
For checking your arithmetic, yes. If your cost per support conversation is thirty times Anthropic's published $0.0037, that's worth investigating. For judging whether your total is acceptable, no, because they price a scenario the vendor chose rather than the one you run.
How do I know if my spend is a problem?
When cost per unit sold rises while your price holds, or when gross margin falls quarter on quarter. Both are answerable from your own data and neither needs anyone else's. A merely large total proves nothing by itself.
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.
How Culpa works
Find the culprit. Not just the total.
Your dashboard shows what you spent. It stops short of who spent it. Culpa shows the conversation, the user and the feature behind it.
Your prompts stay local.
Culpa runs on your own infrastructure. What you send to a model reaches us at no point.
Every dollar has a name.
Follow any charge to the conversation, the user, the feature and the customer behind it.
See the bill before it lands.
Cost your next feature before you ship it. You get the likely bill and the worst case, at best, median, p90 and p99.
Three steps to your first answer.
Change one base URL.
Or drop in the Python or TypeScript library.
Find your most expensive conversation.
In the first session, not the first week.
Cost your next feature before you ship it.
Why the bill went up
Example dashboardCalls traced
418,209
across 3 projects
Spend this week
$378.41
+ $182 vs last week
Failed calls
312
74% retried, and you paid for all of them
+ $182 this week traced to one culprit
Spend over 14 days
Most expensive users
Next week forecast
Graded against reality. Accuracy shown as results land.
Keep reading
Sources: Anthropic pricing. Last reviewed 2026-08-05, rates effective 2026-07-02. Plain text version.