01 · Context Why this exists
Three dated provider events landed on agent operators in one month:
- Anthropic retired claude-opus-4-1 on 2026-08-05. Agents pinned to it don't
degrade politely — the API says no. This dataset was measured on
2026-08-04, the retirement eve — the "before" side of that swap can no longer
be re-measured.
- Claude Sonnet 5's introductory pricing ($2/$10 per MTok) ends 2026-08-31;
standard pricing is $3/$15 from September 1 (provider price list, checked 2026-08-04).
- Claude models from Opus 4.7 onward use a new tokenizer documented as producing
"approximately 30% more tokens for the same text," workload-dependent.
Plenty of commentary repeats the "~30%" figure. Nobody had published measured numbers
for realistic agent configurations — system prompt + tool schemas + user turns
— let alone for German and Czech. So we measured.
02 · Results Headline results
Input tokens; each cell is the median of per-bundle deltas across 8 agent
configurations per language, with the min…max range.
03 · Findings What surprised us
Our own hypothesis was wrong, and we're publishing that. We pre-registered the
expectation that Czech — diacritics, rich morphology — would inflate most under
the new tokenizer. The opposite is true: Czech inflates least in every pair, and got outright cheaper in the Opus 4.6 → 4.8
swap (median −3.0%). The new tokenizer family handles Czech better than the old one did. If
you operate Czech-language agents, the tokenizer change is the smallest of your
migration worries.
The "~30% more expensive" shorthand overstates the Sonnet swap for real agent
configs. On realistic bundles we measure +5.6% to +11.2% (median, by
language) for claude-sonnet-4-6 → claude-sonnet-5 — well under the documented upper
band, which is honest of the provider (they say "up to"; commentary tends to drop the
qualifier).
German consistently inflates more than English in the Sonnet and Opus 4.6→4.8
pairs — a data point that matters if your clients are DACH businesses.
04 · Prices The money axis
Token deltas are only half the bill. Per the provider's price list on the measurement
date:
- P1 (forced swap): claude-opus-4-1 was $15/$75 per MTok in/out;
claude-opus-5 is $5/$25 — per-token prices fall by two-thirds on both axes. Our
measured input inflation (+10.5% to +21.4%) offsets only a fraction of that: input-token bills fall roughly 60% at list prices for the same traffic. Retirement forced your hand, but on the input side it forced it
downhill.
- P2 (Sonnet): two separate effects, don't conflate them. (a) Continuing
Sonnet 5 users: on September 1 the intro discount ends — +50% per token, no token change. (b) Migrating from Sonnet 4.6 after September 1 (price parity at $3/$15):
the bill delta is the token delta — +5.6% to +11.2%. Migrating during
the intro window is net cheaper than staying on 4.6.
- P3: claude-opus-4-6 and claude-opus-4-8 share a price ($5/$25) — the bill delta
equals the token delta, including the Czech decrease.
Boundary, stated plainly: this dataset measures the input side only.
Output-token volume is workload-dependent and unmeasured here — and claude-opus-5 has
thinking enabled by default, which adds output tokens a static count cannot predict.
Measuring what a swap does to a running agent — cost, parameters, behavior
consistency — is per-agent re-verification work, which is the paid desk, not this
dataset.
05 · Method Method
- Corpus: 24 synthetic config bundles — 8 agent archetypes (e-commerce support,
booking, voice agent, invoicing assistant, lead qualification, internal docs Q&A, ERP
order desk, hospitality FAQ) × 3 languages (EN/DE/CZ). Each bundle: system prompt, 3–4
tool schemas, 5 representative user turns. No client data. The corpus is published in
full below — judge its representativeness yourself.
- Measurement: the provider's
count_tokens endpoint, per (bundle ×
model), N=3 identical calls. All 144 measured cells were run-to-run identical
(the endpoint is deterministic; N=3 documents that rather than assumes it).
- Sealing: the raw results file is committed to by SHA-256 (below).
- Dating: measured 2026-08-04. Tokenizers don't drift daily, but every
claim on this page is a claim about that date.
06 · Verification Parameter changes, live-verified
Documentation claims about breaking parameter changes, verified against the live API on
the measurement date:
07 · Data Full distribution
No cherry-picking: every bundle, every pair, and the underlying input-token counts
(median of N=3 runs per cell).
08 · Reproducibility Reproduce it
sha-256 · raw results file 2138a6229eb5cd3600a8c7b6529d7367cb9073b35b62b03a72cd8eaf3494d82f
Method questions, corrections, or a swap you want measured:
audit@strattlabs.com. We hold our own
instrument to the standard we hold anyone's: measurements, never verdicts; when our
instrument errs, we withdraw the finding in writing. See
how we test.