Jev tokens: what a request is billed, measured
TypeSafe publishes the price per token but not how a request turns into tokens. We measured it: 479 billed requests to jev-1.13 on 2026-09-23, open code and data. Paste a request below to see what it costs.
- 260 tokens fixed overhead on every request
- +8 per question, plus ~1 per English instruction word
- +8 per Choice option (~21 with a short description)
- +8 per Score level
- English 4.92 chars/token; Chinese and Japanese ~1
- Limits: 32,768 state + longest question, 65,536 total (overhead excluded)
Paste a System One request
- Fixed per-request overhead
- 260
- State (English)
- 23
is_billingnoul · 8 instruction tokens- 16
departmentchoice · 6 instruction tokens, 4 options- 89
urgencyscore · 5 instruction tokens, 4 levels- 50
Fits: state plus the longest question is 112 of 32,768 tokens.
An estimate from measured coefficients: median error 1.4% on 365 billed requests, worst when question text is not in English. Cost at $0.042 per 1M input tokens (docs.typesafe.ai, checked 2026-09-23); output is free. Computed in your browser; nothing is sent.
Characters per token, and whether answers move
| Language | Chars/token | Tokens, EN questions | Tokens, translated | Noul shift (EN q / translated) |
|---|---|---|---|---|
| English | 4.92 | 609 | — | baseline |
| Chinese | 1 | 624 | 695 +11% | 0.017 / 0.018 |
| Japanese | 1.01 | 645 | 820 +27% | 0.017 / 0.023 |
| Korean | 1.44 | 632 | 728 +15% | 0.015 / 0.042 |
| Spanish | 4.44 | 617 | 675 +9% | 0.013 / 0.026 |
| Hindi | 1.72 | 663 | 980 +48% | 0.010 / 0.031 |
| Arabic | 1.49 | 646 | 905 +40% | 0.016 / 0.035 |
| Russian | 1.78 | 665 | 916 +38% | 0.016 / 0.067 |
Keep questions in English even when the text is not. Translating the questions as well cost 9–48% more tokens and, in all seven languages, moved yes/no answers further from the English baseline than translating the ticket alone.
Where requests stop fitting
We found both limits by bisection on real requests. With one short question, a state of 32,688 content tokens passed (billed 32,948, overhead included) and one of about 32,813 returned HTTP 400. With a 20,000-token state and many questions, 65,388 content tokens passed (billed 65,648) and about 65,644 did not.
Both brackets contain the binary limits 32,768 (32K) and 65,536 (64K), and both passing requests billed more than the limit, so the ~260-token overhead does not count toward either. TypeSafe documents the same limits as 32k and 64k.
There was no cap on the number of questions: one call with 1,000 questions succeeded. Measured through OpenRouter, which lists a 32K context; TypeSafe documents 64K, and the 64K total held here.
What changes the answer, and what does not
| Change | Mean shift | vs repeat noise |
|---|---|---|
| 120 unrelated questions in the same call (Noul) | 0.002 | 0.7× |
| Choice options reordered (Choice) | 0.020 | 2.2× |
| Yes/no question paraphrased (Noul) | 0.024 | 8.0× |
| Instruction as a JSON object (Noul) | 0.028 | 9.3× |
| Choice option descriptions removed (Choice) | 0.096 | 10.7× |
Noul: absolute change in probability. Choice: total-variation distance between the two probability distributions. No change above altered a Choice answer’s top option.
Batch freely. Asking 16, 60 or 120 unrelated questions in the same call left the target answers exactly as noisy as asking them alone, and latency stayed flat: 426 ms at 4 questions, 436 ms at 124. What batching saves.
Treat wording as part of the model. Option order barely matters, but descriptions and phrasing move probabilities by a few points. If your code acts on a threshold, re-check it after any rewording.
Confidence tells you which answers are fragile. Across 960 answers we reworded or translated, the lowest-confidence third of Choice answers moved 15.6× as much as the highest third (Spearman ρ -0.79), and all 5 changes of top option were in that low third. For Score answers the gap was 7.5×. Acting only above a confidence threshold, and sending the rest to review or a fallback, is what TypeSafe’s confidence-gated routing pattern recommends; this is measured evidence for it. Analysis.
Jev token questions
How many tokens does a Jev request cost?
About 260 input tokens of fixed overhead, plus the state, plus each question: about 8 tokens of framing, 1 per English instruction word, and 8 per Choice option or Score level. Measured on 2026-09-23 over 479 billed requests to jev-1.13 via OpenRouter.
How many characters is one Jev token?
About 4.92 characters of English prose and 4.44 of Spanish, but only about 1 of Chinese or Japanese, 1.44 of Korean, 1.49 of Arabic, 1.72 of Hindi and 1.78 of Russian. A JSON object as state bills about 2.4 characters per token.
Where do Jev's context limits cut off in practice?
TypeSafe documents 32k for state plus the longest question and 64k in total. In our tests via OpenRouter, a request with 32,688 content tokens of state plus its longest question passed and one of about 32,813 was refused; in total, 65,388 passed and about 65,644 was refused. That brackets binary limits of 32,768 and 65,536, and the ~260-token overhead does not count toward either.
Is there a limit on how many questions one Jev call can ask?
Not one we could find: a single call with 1,000 questions succeeded. The token limits bind first.
Should I translate my questions if my text is not in English?
Keep them in English. In our test, translating the questions as well as the ticket moved yes/no answers further from the English baseline than translating the ticket alone, in all seven languages, and cost 9–48% more tokens (48% for Hindi).
Does Jev's confidence tell me when an answer is unreliable?
It tells you which answers are fragile. Across 960 answers we reworded or translated, the lowest-confidence third of Choice answers (confidence 0.54–0.92) moved 15.6 times as much as the highest third, and every change of top option happened in that low third. Gate on confidence: route low-confidence answers to review or a fallback.
Does rewording a Jev question change the answer?
It can shift the probabilities, not usually the top option. Reordering Choice options never changed the top option in our test; removing option descriptions or rephrasing a yes/no question moved probabilities by 2–10 points on average. If you act on a probability threshold, re-check it after rewording.
Method, code and data: jev-fanout-bench, round 2. 479 billed requests (plus 8 deliberate over-limit probes) to jev-1.13-20260917 through OpenRouter’s TypeSafe-compatible endpoint (provider TypeSafe), $0.036 billed. Answers were used only to compute the differences above and are not published. Not affiliated with TypeSafe AI.