The bill went up, the price didn't

Somewhere on your team, someone checked Anthropic's pricing page, confirmed the per-token rate hadn't changed, and closed the tab satisfied nothing was wrong. Then the invoice landed 20-40% higher than last month. No new features shipped. No traffic spike. Same prompts, same model tier, same published price per million tokens. The math should hold. It doesn't — because the thing you're paying for isn't a stable unit of text. It's a token, and Anthropic quietly changed how many tokens your text turns into.

This is the story behind a growing number of confused finance tickets this month: teams that migrated to newer Claude models expecting flat or lower costs, and got a real cost increase with zero code changes and zero pricing changes on paper.

What actually changed: a new tokenizer under the hood

Anthropic's documentation now confirms that Opus 4.7 and later, Sonnet 5, Fable 5, and Mythos 5 all run on a new tokenizer — the component that splits your raw text into the discrete units (tokens) a model actually processes and gets billed on. Older Claude models used a different tokenizer that packed more characters into each token. The new one is documented to produce roughly 30%, and in some cases up to 35%, more tokens for identical input text.

That's not a pricing change. It's a unit-of-measure change. Imagine your electricity company kept the price per kilowatt-hour identical but redefined a kilowatt-hour to be 30% smaller. Your rate didn't move. Your bill did.

The numbers: 30-35% baseline, up to 73% on real code

Anthropic's own figures describe the general-text impact. Independent benchmarking published this week goes further, specifically on code — arguably the highest-volume, highest-cost workload for most AI-using engineering teams. On identical code samples, researchers found Claude's new tokenizer generating up to 73% more tokens than GPT-5.6 needs to represent the same content.

If your team runs an AI coding assistant, a code-review bot, or an autonomous coding agent against Claude, this isn't a rounding error — it's the difference between a task costing $0.40 and $0.69 with the exact same prompt, the exact same output quality, and the exact same price-per-token on the invoice.

Why $/million-token pricing tables quietly lie

Every pricing comparison chart you've ever seen — including the ones on provider websites — treats a token as a fixed, comparable unit across models and vendors. It isn't. Tokenization is model-specific. GPT-5.6, Gemini, and Claude all split the same sentence into a different number of pieces, and each model family's tokenizer can change between versions, as Anthropic just demonstrated.

This means the honest question is never "which model is cheaper per token?" It's "which model is cheaper per completed request, task, or outcome?" A model that charges half as much per token but needs twice as many tokens to say the same thing is not cheaper. It's identical — or worse, if its tokenizer is less efficient on your specific content (code, JSON, non-English text, and structured data all tokenize very differently across providers).

The Sonnet 5 trap: intro pricing expires August 31, 2026

Here's the part most teams haven't noticed yet. Sonnet 5's current introductory pricing is masking the tokenizer effect for early adopters — the promotional rate is low enough that even with 30% more tokens per response, the total bill looks flat or lower than the previous model generation. That introductory pricing ends August 31, 2026.

After that date, teams that migrated to Sonnet 5 assuming they'd locked in a permanent cost improvement will see two effects stack at once: the promotional discount disappearing and the tokenizer's ~30% token inflation becoming fully visible in the bill for the first time. If you migrated workloads to Sonnet 5 this year, the number you're looking at in your dashboard today is not the number you'll see in September. Budget for it now, not after the invoice arrives.

It's not just Anthropic

This is a structural problem with how AI pricing gets compared, not an Anthropic-specific failure. GPT-5.6, Gemini, and every other frontier model use their own tokenizers, and those tokenizers change across versions without warning or fanfare — usually documented, if at all, in a changelog most finance teams never read. Every time a provider ships a new tokenizer, every static price comparison built on $/million-tokens becomes stale, even though nothing on the pricing page moved.

The practical implication: any cross-provider comparison based on published rates alone is already out of date by the time you read it. The only reliable comparison is one built from your own usage data, measured after the fact, per model version.

The fix: measure cost-per-task, not cost-per-token

Stop comparing rate cards. Start comparing outcomes. For any workload you run across multiple models or providers, track:

• Cost per completed request — total spend divided by successful responses, not tokens processed.
• Cost per resolved task — for agents or multi-step workflows, the full chain of calls needed to finish one job, including retries.
• Token-to-output ratio by model version — how many tokens each model burns to produce a comparable output, tracked over time so a tokenizer change shows up immediately instead of three weeks later in accounting.

This applies whether you're running customer support automation, a coding agent, or B2B outbound at scale — an autonomous prospecting agent like AutoReach that generates qualified leads for pennies each is still an AI workload with real per-lead API cost, and that cost only makes sense measured per outcome, not per token.

How to audit your workload before/after a model migration

1. Pull actual token counts (input + output) per request from your provider's usage logs for a two-week window before and after the migration.
2. Calculate real cost per request, not the advertised rate — divide total spend for that workload by number of completed requests.
3. Compare the token-count delta on identical or near-identical prompts run through both model versions.
4. Flag any workload where token count per request jumped more than 15% — that's your tokenizer-change canary.
5. Re-run this audit after any provider announces a tokenizer or model update, and again before any introductory pricing period ends.

If a workload is now consistently expensive per task, it may be a candidate for a smaller, purpose-built model instead of a general frontier model — fine-tuning an open-weight model on your own data via a service like InfoPlatform.ai and owning the weights can sidestep both the tokenizer volatility and the rate-card unpredictability of closed APIs entirely.

Why a live dashboard beats a static pricing table

Pricing pages are snapshots. Tokenizers, discount windows, and model versions change continuously, and none of that shows up until the invoice does. The only durable defense is watching real, normalized spend across every provider you use, continuously, so a 30% token inflation or an expiring promotional rate shows up as a trend the week it starts — not a surprise the month it fully lands. That's the entire premise behind AICosts.ai: it ingests actual usage and billing from Anthropic, OpenAI, Google, and 50+ other providers, normalizes it into comparable cost-per-request and cost-per-task metrics, and flags spend anomalies like this one before finance finds them in the invoice.

FAQ

Why does Claude cost more than GPT for the same prompt?

Not because Anthropic's rate is higher — often it isn't. It's because Claude's tokenizer can split the same text into significantly more tokens than GPT's tokenizer, especially for code, where independent benchmarks show up to 73% more tokens on identical samples. More tokens at the same per-token price means a higher total cost for the same task.

Does the tokenizer change affect Opus 4.8 tokens vs Opus 4.6?

Yes. Opus 4.7 and later versions run on the new tokenizer, so any Opus version from 4.7 onward will consume roughly 30-35% more tokens than Opus 4.6 and earlier for equivalent text, even with identical prompts and outputs.

What happens when Sonnet 5's intro pricing ends August 31, 2026?

Two cost increases hit simultaneously: the promotional discount rate expires, and the tokenizer's roughly 30% token inflation — currently masked by that discount — becomes fully visible in your bill. Model this now rather than reacting to the September invoice.

How do I compare AI pricing honestly when tokenizers differ?

Ignore $/million-token rate cards for cross-provider comparisons. Measure real cost per completed request or per finished task using your own usage logs, for the exact workload you run, and re-check it whenever a provider ships a new model version or tokenizer.