VynarisEarly betaGet your API key

Blog · 2026-07-15 · Vynaris Team

Anthropic's new tokenizer: same sticker, ~30% more tokens per request

Anthropic's new tokenizer emits ~30% more tokens for the same text, so Opus 4.8 costs ~30% more than 4.5 at an identical sticker. Verified 2026-07-15.

Anthropic's newer models kept their predecessors' sticker prices but changed the tokenizer underneath them. The pricing page states the new tokenizer "produces approximately 30% more tokens for the same text" (verified 2026-07-15). That makes Claude Opus 4.8, at the same $5/$25 per million tokens as Opus 4.5, cost about 30% more for an identical request. Same price list, higher bill.

TL;DR

What changed and why it matters

Model comparisons run on price per million tokens. That number assumes a token is a fixed unit of text. It is not. A tokenizer decides how many tokens a given string becomes, and Anthropic shipped a new one with its recent models. From the pricing page, verbatim:

Claude Opus 4.7 and later Opus models, Claude Fable 5, Claude Mythos 5, Claude Mythos Preview, and Claude Sonnet 5 use a newer tokenizer that contributes to their improved performance on a wide range of tasks. This tokenizer produces approximately 30% more tokens for the same text. The exact increase depends on the content and workload shape. Claude Sonnet 4.6 and earlier models use the previous tokenizer.

Read that as a hidden price change. When the token count for the same text rises 30% and the price per token holds, the cost per token of a fixed piece of work rises 30% too. The sticker did not move, so a spec-sheet comparison misses it entirely. The only place the change is disclosed is a footnote under the pricing table.

Which models are on which tokenizer

The boundary runs mid-generation, which is what makes it easy to miss:

Tokenizer               Models
----------------------  -----------------------------------------------------------------
New (~30% more tokens)  Opus 4.7, Opus 4.8, Fable 5, Mythos 5, Mythos Preview, Sonnet 5
Previous                Opus 4.6, Opus 4.5 and earlier, Sonnet 4.6, Sonnet 4.5, Haiku 4.5

Two adjacent Opus versions, 4.6 and 4.8, share the exact same $5/$25 sticker and sit on opposite sides of the tokenizer line. Upgrading from 4.6 to 4.8 for the quality gain also raises the bill on the same prompts, with nothing on the price list to warn you.

Effective price per unit of text

To compare cost per unit of text rather than per token, scale the new-tokenizer models' per-token price by the inflation factor. At Anthropic's ~30% (factor 1.30), the effective prices per million tokens of the same text become:

Model             Sticker in/out  Tokenizer  Effective in/out (per text)
----------------  --------------  ---------  ---------------------------
Opus 4.5          $5 / $25        old        $5.00 / $25.00
Opus 4.8          $5 / $25        new        $6.50 / $32.50
Sonnet 4.6        $3 / $15        old        $3.00 / $15.00
Sonnet 5 (intro)  $2 / $10        new        $2.60 / $13.00
Sonnet 5 (Sep 1)  $3 / $15        new        $3.90 / $19.50
Fable 5           $10 / $50       new        $13.00 / $65.00

"Effective" here means the price you actually pay to process a fixed document, since that document turns into ~30% more tokens on a new-tokenizer model. The output side inflates the same way: to express the same answer, the model emits ~30% more output tokens.

A worked request

Take one representative task: summarize a document that tokenizes to 10,000 input and 800 output tokens on the old tokenizer. A new-tokenizer model bills the same text and the same answer as roughly 13,000 in and 1,040 out.

Same document, dollars per 1,000 tasks across the tokenizer boundary; new-tokenizer models cost ~30% more at the same sticker
Same document (10k in / 800 out on old tokenizer), $/1,000 tasks. New-tokenizer models billed at 1.30x. Source: Anthropic pricing verified 2026-07-15.
Model             Billed in  Billed out  $/task   $/1,000 tasks
----------------  ---------  ----------  -------  -------------
Opus 4.5          10,000     800         $0.0700  $70.00
Opus 4.8          13,000     1,040       $0.0910  $91.00
Sonnet 4.6        10,000     800         $0.0420  $42.00
Sonnet 5 (intro)  13,000     1,040       $0.0364  $36.40
Sonnet 5 (Sep 1)  13,000     1,040       $0.0546  $54.60

Opus 4.8 costs $21 more per 1,000 of these tasks than Opus 4.5, a 30% increase from a tokenizer footnote. Put your own token counts through the calculator to see the effect on your workload shape.

The Sonnet 5 trap on September 1

Sonnet 5's math is the one most likely to surprise a team mid-quarter. Today, at the $2/$10 introductory rate, Sonnet 5 is genuinely cheaper than Sonnet 4.6 even after the token inflation: $36.40 versus $42.00 per 1,000 tasks, about 13% less. Teams benchmarking this summer will bank that saving.

The introductory pricing ends 2026-08-31. On 2026-09-01 Sonnet 5 moves to the standard $3/$15, matching Sonnet 4.6's sticker exactly. But because Sonnet 5 sits on the new tokenizer, the same workload then costs $54.60 versus Sonnet 4.6's $42.00, about 30% more for the same text. The sticker will read "identical to 4.6." The bill will not. Anyone who standardized on Sonnet 5 during the promo should model the September step now, not in the August invoice. We work the raw before-and-after in the Sonnet 5 introductory-pricing analysis; the tokenizer is the multiplier that makes the real jump larger than the 50% sticker rise implies.

What it means for model choice

Two practical consequences.

First, cross-tokenizer per-token comparisons are apples to oranges. "Model X is $5/MTok and model Y is $5/MTok, so they cost the same" is false when one uses the new tokenizer. This also reaches across vendors: comparing Opus 4.8 to a competitor on price per million tokens understates Opus by whatever its inflation is, because the two count the same text differently. The only exact comparison is to bill your real task on each and read the invoices, which is the discipline behind our July 2026 cost-per-task tables.

Second, the inflation interacts with every other multiplier. Prompt caching, the Batch API, and long-context billing all operate on token counts, so a 30% larger token count scales those line items too. For a coding agent that re-sends a large prefix each turn, the inflated input token count compounds across the whole loop, a pattern we priced in the coding-agent cost model.

Where we are wrong

The honest caveats matter here, because the finding cuts against Anthropic and we are not going to overstate it.

The 30% is approximate and workload-dependent by Anthropic's own words. Code, non-English text, and heavy punctuation tokenize differently, so your figure could be 20% or 40%. Do not hardcode 30%. Measure it: run the same corpus through the count_tokens endpoint on a 4.6-era model and a 4.8-era model and take the ratio. That is the only number that is yours.

The tokenizer change is not purely a cost story. Anthropic states it "contributes to their improved performance," so more tokens can mean better output per task, which can lower total cost if it cuts retries or lets a cheaper model succeed. And generational price cuts have run the other way and larger: Opus 4.1 was $15/MTok input, and Opus 4.5 onward is $5, a 3x drop that dwarfs a 30% tokenizer effect. The point is not that new Claude models are a bad deal. It is that the price list alone no longer tells you the per-request cost, and you have to measure across the tokenizer boundary to compare honestly.

Sources

Prices and the tokenizer note are current as of 2026-07-15. We re-verify monthly and stamp updates, and Sonnet 5's step to standard pricing on 2026-09-01 will get an "Updated" banner.

Want per-request costs that count real tokens, not sticker prices, across every Claude model and its competitors? Vynaris is an OpenAI-compatible gateway that shows the true per-request cost and routes to the cheapest right-sized model. One base URL swap. Get an API key at vynaris.com.