The silent token inflation.

Simon Willison, in a brief note dated 20 April 2026, flags a fact most teams running Claude in production haven't measured yet: the new Opus 4.7 tokenizer counts the same text differently. Measurements in hand, the same system prompt jumps from five thousand to seven thousand tokens. At the same price per token, that's a forty percent operational increase disguised as a model update.

The data, in brief

Anthropic has announced that the new tokenizer introduces a multiplier between 1.0 and 1.35 depending on content type. Willison's checks, conducted with a public tool that compares models on the same input, return a less reassuring observation: on a real system prompt, the multiplier turned out to be 1.46. Text-dense PDFs behave better, around 1.08. High-resolution images blow up, but in that case the extra consumption is justified by greater analytical capability. The practical conclusion is simple and unpleasant: upgrading to Opus 4.7 without recalibrating prompts means paying a higher bill for the same result.

An arithmetic example to grasp the order of magnitude. A well-instrumented inbound-lead qualification agent runs about 30,000 calls per month, with an average system prompt of 4,500 tokens on the old tokenizer. Monthly input cost on Opus 4.7 at current list price is roughly: 30,000 calls × 4,500 tokens × input price per million tokens. Upgrading to the new tokenizer, the same prompt is worth 6,300 tokens (1.4 average multiplier on Italian text). Same functional outcome, same list price, but monthly input cost up forty percent. On an annual scale, several thousand euros that vanish from margin without anyone signing an increase.

Why it matters more than it seems

The price per million tokens is unchanged. That's the clever detail. The public narrative remains "same cost, better model", but the real cost per request rises, because the same sentence produces more tokens. For anyone running an agent in production at serious volumes — outreach, classification, extraction, support — margin erodes without any list changing. It's silent inflation: you don't see it in the quote, you see it in the monthly bill two cycles later.

The second layer of the problem is less technical and more organisational. Most companies that adopted Claude in the past twelve months don't have an internal function monitoring cost per request, average tokens per call, or monthly drift of these metrics. When a provider changes tokenizer, these companies find out from the accountant, not from the logs. It's the same recklessness with which cloud contracts were signed in 2018, multiplied by the speed at which models update today.

The three prompt types most hit

Not all prompts suffer equally. Having audited dozens of stacks in recent months, we see three clear categories of prompts that pay most for a tokenizer change.

1 · Long, redundant system prompts. Typical of recently written or inherited agents: three paragraphs of "you are a professional assistant that…", four few-shot examples no longer needed now that models are more capable, a list of constraints repeated in three different forms. They're the prompts that bloat by inertia, not by necessity. The multiplier here is worst — around 1.4-1.5 — because didactic text contains many common words with less-than-average tokenisation efficiency.

2 · Documents dense with sector terminology. Technical manuals, contracts, clinical documentation, regulations. Sector terminology (long Latin roots, acronyms, technical proper names) segments into more tokens than common words. Typically multiplier 1.25-1.35.

3 · Long multi-turn conversations. Each turn adds context. If the tokenizer counts each message slightly more than the previous, the effect compounds. A conversation that at turn 1 cost X by turn 12 costs 1.4X just from accumulation, before the new response is even generated. For agents that maintain extended conversational memory, it's the worst case.

Safe prompts, by contrast, are the short and direct ones on text-dense PDFs, where the multiplier stays close to 1.05-1.08. They're often structured-extraction tasks on documents already tokenised efficiently.

Anonymised real case · before and after

An outreach agent in production at one of our clients, active for seven months on Opus 4.6, generated personalised emails from a system prompt of 5,200 tokens (target customer description, voice rules, few-shot examples, legal constraints on what the agent can promise). On a volume of 18,000 emails/month, monthly input cost roughly five hundred thirty-two euros.

Upgrade to Opus 4.7 without recalibration: same prompt, same volume, monthly input cost rose to seven hundred forty-five euros. The client only saw it after two billing cycles, on weekly charts the vendor had never shown. Extended annual difference: over two thousand five hundred euros on a single agent, with a mid-size client.

What we did in half a day of work. Read the system prompt line by line. Cut two paragraphs of redundant didactic instructions (the agent wasn't new, it didn't need to be "reminded"). Removed three of four few-shot examples (Opus 4.7 no longer requires them for the same quality). Compacted the voice rules from six sentences to two. Final prompt: 2,900 tokens against the original 5,200. Monthly cost dropped to four hundred eighteen euros, below the pre-upgrade baseline. Same output quality, measured on a sample of one hundred emails compared double-blind.

Operational moral: a model upgrade is the right moment to revisit the prompt, not to ignore it. Most production prompts are bloated for historical reasons the new model renders obsolete.

Invisible inflation isn't only Anthropic

It would be comforting if this were a single-provider problem. It isn't. The same pattern we've seen on historic OpenAI transitions (tokenizer changes across major versions) and Google (modification of the effective context window on Gemini 3). Every provider, every two-four releases, adjusts something that moves consumption without moving the list price. The public narrative always remains "same price, better model". Real task cost moves by a non-zero factor.

For those running a multi-provider system — the Jigen operational pattern — the consequence is double. On one hand, token inflation can hit any component. On the other, a re-routing option always exists: if Opus inflates by forty percent on a task, and Sonnet 4.6 does the same task at a cost that, even with its own multiplier, remains lower, the flow shifts. For those tied to a single provider, this escape route simply doesn't exist — they pay the inflation and that's it.

Jigen's position

Every project that enters our twenty-one days is instrumented, from week one, with three mandatory numbers: average tokens per request, cost per useful outcome, and percentage of calls where the system prompt weighs more than the user message. They're banal metrics, yet most stacks seen in audit don't keep them. Eyes are on the result, not on consumption. It's exactly the mistake that makes AI agents fragile in production: they work until you upgrade the model, then they stop working at acceptable margin.

The Opus 4.7 upgrade is a good discipline drill. Whoever keeps their prompts under version control, with token logs per call, can respond to the increase in two ways: recalibrate the system prompt — often bloated by habit, not necessity — or change model on the part of the flow that doesn't require Opus. Whoever keeps nothing under control pays the forty percent and waits for the next upgrade.

How to recalibrate a system prompt in 30 minutes

A practical procedure for those wanting to act on their stack immediately. Works regardless of provider — applicable to OpenAI, Anthropic, Google in the same way.

Minutes 0-5 · measure the baseline. Count the tokens of the current system prompt on the tokenizer of the current model. Most providers expose an endpoint or public tool. Note the number. That's your starting point.

Minutes 5-15 · cut the didacticism. Look for these three patterns and delete them without mercy: (a) sentences explaining to the model what it is ("you are an assistant"), (b) repetitions in different forms of the same constraint, (c) pre-2025 few-shot examples the current model no longer requires for the same quality. Typically, on a 5,000-token prompt, between 1,500 and 2,500 tokens are cut with no measurable quality loss.

Minutes 15-22 · replace common words with dense ones. The tokenizer counts fewer tokens on common words and standard technical roots. "Please evaluate carefully" becomes "evaluate". "The system must ensure that" becomes "ensure that". Ten or fifteen substitutions on a system prompt are worth a further 200-400 tokens.

Minutes 22-28 · validate the output. Run the shortened prompt on a sample of thirty to fifty real inputs. Compare outputs with those from the original prompt. If quality is identical (operational definition: the human reviewer can't tell which is which), the shortened prompt goes to production. If it degrades, restore the necessary pieces one at a time.

Minutes 28-30 · commit, log, monitor. The new prompt is versioned in the repository, with commit message declaring the delta in tokens. An internal dashboard records consumption per call before and after. The saving becomes measurable in the first week.

Thirty minutes. Often the saving exceeds the value of half a day's work within the first month.

Quick test · are you safe from token inflation?

A diagnostic grid for the leadership of a company wanting to know if its AI agent is well managed. Five questions:

1. Do you know, today, how many tokens on average your agent consumes per request? Healthy answer: a precise figure (e.g. "4,500 input tokens average, 350 output"). Suspicious answer: "we don't know, it depends".
2. When the provider updates the model, does anyone measure consumption before and after? Healthy: yes, always, with a written report. Suspicious: "we check it still works".
3. Is the system prompt under version control (git or equivalent)? Healthy: yes, every change has a commit. Suspicious: "it lives in a configuration field we change at runtime".
4. Is there a weekly chart of cost per request? Healthy: internal dashboard or vendor report. Suspicious: "you see it in the monthly statement".
5. Have you identified the most expensive flow in your stack and you know why? Healthy: yes, it's X, costs Y, because Z. Suspicious: "all flows are similar".

Three yes out of five: you're managing it well. Two or fewer: there's inflation eroding margin without you knowing. The first week to measure, the second to recalibrate, the third to stabilise. Three weeks — coincidence — of the Jigen cycle.

The implication for those who lead

For a CEO or founder integrating AI into production, the point isn't to understand how a tokenizer works. It's to demand their operations team, internal or external, show three things by next Monday: the chart of cost per request over the last four weeks, the list of active system prompts with their token weight, and identification of today's most expensive flow. If those three outputs don't exist, the problem isn't Anthropic. The problem is that an AI budget is being spent without accounting.

The operational advice is specific. Don't upgrade models in production without a before/after measurement. Don't assume that price per token coincides with price per request. Keep a counting tool — Willison's public one is a good starting point — hooked to the most trafficked system prompt, and run it every time a model changes version. It's a half-day investment that protects margin decimals over the next twelve months.

Willison's datum is small but telling. Enterprise AI isn't managed with the same logic as traditional software, where a minor update is free by definition. Here every vendor release is a cost negotiation. Whoever understands it in time pays what they see. The others discover the bill at year-end close.

Source: Simon Willison, "Claude Token Counter, now with model comparisons", simonwillison.net, 20 April 2026 — read the original ↗. Linked tool: claude-token-counter ↗. Jigen reading: Willison's observation isn't an isolated case — it's the signal of a dynamic every AI provider repeats every two-four releases. The operational answer isn't to demonise a vendor, it's to instrument the stack: average tokens per call, cost per useful outcome, versioned prompt control, weekly dashboard. Whoever doesn't do it signs a blank check for the next update. Multi-provider and prompts under control are the two levers that neutralise silent inflation.