Guides / llm tokenizer change cost
When a tokenizer change raises your bill and the rate card doesn't move
A tokenizer change alters how much text becomes one billable token, so identical prompts cost different amounts at an identical published rate. Anthropic states Claude 4.7 and later produce about 30% more tokens for the same text. Culpa, a local-first LLM cost, margin, and forecast ledger, meters the tokens each call actually reported rather than the rate its model advertises.
Why this happens
Every cost comparison people run assumes one thing nobody checks: that a token means the same amount of text on both sides. Anthropic's own pricing page says it doesn't. Claude 4.7 and later models, plus Claude Mythos Preview, use a newer tokenizer that in its words produces approximately 30% more tokens for the same text, with the exact increase depending on content and workload shape. Claude Sonnet 4.6 and earlier use the previous one. The rate card says nothing about this, which is what makes it dangerous. Opus 4.5 through Opus 5 all publish $5.00 and $25.00 per million, so a rate comparison across that group returns a tie, and a bill comparison doesn't. The effect runs the same way on every cross-model comparison anyone does on published rates alone: the newer model looks level and arrives higher. It also runs backwards through your own history, because a month priced before an upgrade and a month priced after aren't measuring the same unit.
What this usually looks like
- You upgraded a model version at an identical rate and the bill moved anyway.
- Two services on the same rate card report very different costs for similar work.
- Your cost per request rose across a model upgrade and the prompt never changed.
- A model comparison run on published rates disagreed with what you were actually charged.
- Your tokens-per-request metric has a step change in it that nobody can explain.
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 |
|---|---|---|
| Comparing two models on rate alone. | A rate is dollars per token and the comparison you want is dollars per request. Those differ when tokenization does. | Send the same input to both and compare the token counts each one reports back. |
| Treating a same-rate version upgrade as cost-neutral. | Across the Claude 4.6 to 4.7 boundary the rate holds and the token count moves by roughly 30%. | Price the upgrade on measured tokens from your own traffic before you roll it out. |
| Comparing months either side of an upgrade. | The unit changed, so a rise can be tokenization rather than demand, and you'll chase the wrong cause. | Record the model version on every call so a step change can be attributed rather than guessed. |
| Assuming the 30% applies evenly. | Anthropic says the exact increase depends on content and workload shape, so a single multiplier is an estimate. | Measure it on your own text rather than applying a headline percentage. |
Run this check tonight
- Send one identical prompt to Claude Sonnet 4.6 and to a 4.7-or-later model, and compare reported input tokens.
- Find your tokens-per-request over the last six months and look for a step rather than a slope.
- Check whether any live model comparison in your planning used rates rather than measured tokens.
- Confirm your ledger records the model version on every call, so a step change is attributable.
The same text, the same rate, two bills
Take a workload of 200M input and 20M output tokens a month as measured on Claude Sonnet 4.6's tokenizer, then move it unchanged to an Opus model on the newer tokenizer. Opus 4.5 through Opus 5 all publish $5.00 and $25.00 per million, so the rate is held constant on purpose. The 30% figure is Anthropic's own, and it applies to the TOKEN COUNT rather than to the price.
A 30% rise in billable tokens is a 30% rise in the bill, and no pricing page announced it. This is the one cost movement that can't be caught by watching rates, which is why it has to be caught by metering tokens.
Every number, with its confidence and source
| Figure | What it means | Confidence | Source |
|---|---|---|---|
| $450.00 per month | modelled increase from a tokenizer change alone, with the published rate held constant | calculated | 200M input and 20M output tokens at $5.00 and $25.00 per million, the published Opus rate from the price book effective 2026-07-02, against the same text at 30% more tokens: $1,500.00 becomes $1,950.00. The 30% figure is Anthropic's own, from platform.claude.com pricing read 2026-08-03. Token volumes are modelled and the 30% is a stated approximation, so treat the shape rather than the figure as transferable. |
What a generic answer can’t know
A rate card is a published fact and a token count is a measurement, and only one of them lives in your own traffic. That's why this particular change is invisible to anything working from provider price lists: the list is still correct, and it was never the thing that moved. Culpa records the model version and the reported token counts on every call, prices them from a versioned price book with effective dates, and keeps the ledger on your infrastructure, so a step change in tokens per request shows up as a fact attached to a version rather than as an unexplained rise. It also means a past month recomputes as it was actually billed, on the tokens that were actually reported, rather than on what today's rate card would suggest.
Questions founders ask next
Which Claude models changed tokenizer?
Anthropic's pricing page states that Claude 4.7 and later models, plus Claude Mythos Preview, use a newer tokenizer, and that Claude Sonnet 4.6 and earlier use the previous one. It says the newer one produces approximately 30% more tokens for the same text, with the exact increase depending on content and workload shape.
Does a tokenizer change show up on the pricing page?
Only as a note. The per-token rates stay exactly where they were, which is what makes this hard to catch. Opus 4.5 through Opus 5 all publish $5.00 and $25.00 per million and sit on both sides of the tokenizer boundary, so a rate comparison across them returns a tie that the invoice contradicts.
How do I measure the effect on my own traffic?
Send one identical prompt to a model either side of the boundary and compare the input token counts each reports back. That number is yours and the 30% is a headline. Anthropic says the increase depends on content and workload shape, so code, prose and structured data won't move by the same amount.
Is a newer tokenizer worse?
Not necessarily, and this page isn't arguing that. Anthropic says the newer tokenizer contributes to improved performance across a range of tasks. The point is narrower: it changes the billable unit, so the upgrade decision needs measuring rather than assuming, and the rate card won't tell you.
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 2026-08-03, rates effective 2026-07-02. Plain text version.