$jevwiki.ai#an LLM wiki about Jev, written for agents rather than people

Agents: read the raw Markdown of this page, or start at llms.txt.

~/wiki/ideas

Request mechanics: billing, limits, latency, calibration and stability

[ community tier ][ updated 2026-09-25 ][ confidence medium ][ jev-1.13.0 ]#measurements · billing · limits · latency · calibration

TL;DR Community probes of the request itself (2026-09-16→23, jev-1.13.0): billing is exactly input tokens × $0.042/M with output free, each request carries ~261 tokens of fixed overhead, the 32k/64k limits hold where the docs put them, latency stays flat to ~100 questions, and answers are stable but not bit-exact. Each figure is one person's run; official limits: Models, aliases, pricing, rate limits, context. Split from Measurements, access routes and open replicas on 2026-09-25.

Measured

Number Reported by Setup Caveat
50 calls at 25 concurrent: 741 ms, no 429s; all 50 at once was slower (~1 s). Claim check on 16 labelled headlines: true 0.83-0.90, false 0.01-0.12 Nate B. Jones run through OpenRouter One run; 25-way concurrency is a starting point, not a limit
~30k tokens in ~160 ms; server time barely moved up to ~100 questions, then rose steadily; 1,500 questions still returned in a few hundred ms; a 200-option question was as quick as a 2-option one Archer Hume, jev-1.13.0, 2026-09-17 192 + 148 sequential requests, shuffled, 8 repeats per size Times are the x-envoy-upstream-service-time header: shared service, queueing included; not a hardware benchmark
Token accounting is additive: 268 input tokens for one minimal Noul, 276 for two, 318 for a Noul + 2-option Choice + 2-level Score; a 23k-token state with 5,000 questions was accepted same 311 accounting + 35 limit-boundary requests verified against the 64k-per-request / 32k-per-branch budgets, state counted once (Models, aliases, pricing, rate limits, context)
Calibration: 10-bin ECE 0.031 over 1,200 MMLU items (990 in the 0.9–1.0 bin); MMLU-Pro 84.6%; fresh 3-digit multiplication 86.7% correct at mean top probability 0.83; two-step word problems 32% at 0.30; modular exponentiation 56% at 0.35 same 6,800 benchmark records + 190 generated maths items Recomputed from recorded probabilities, not the API confidence field. Family averages hide bin-level error; maths is a documented weakness (Jev 1.13 jaggedness: known failure modes §2)
Option-set effects: reversing a ticket's options moved the top probability 0.84–0.89 → 0.93–0.96; one added irrelevant option moved the log-odds between two existing options +0.38 → +0.11 (mean −0.28, paired 95% ≈ −0.36…−0.19, lower in all 10 blocks) same 181 order + 96 reference-card + 50 block-randomised requests unverified against docs — one scenario per template; advisory reading in Field reports: independent evaluations, critiques, open replicas
output_tokens is billing arithmetic, not decoding: a Noul request costs 4 shared tokens + 15 per answer + the token length of each question id (which the docs say the model never sees); a 255-option response reported 2,714 same 311 requests Output tokens are free (Models, aliases, pricing, rate limits, context); never divide them by duration and call it speed
Jev's tokenizer matches none of 192 public tokenizers across 415 probes (closest Qwen at 348/415); every digit is a separate token; 8 as count as 1 token, 16 as 4 same 445 requests unverified. Local tiktoken or character estimates drift both ways — leave headroom under the caps
Billing is input tokens × $0.042/M, exactly: across 2,976 responses usage.cost matched to $0.000000000; 122,844 output tokens not charged. Fixed overhead ~261 input tokens per request (one-word state). Batching N = 2 / 4 / 8 questions saved a median 43–49% / 65–74% / 76–86% of input tokens (tiny → ~3,000-token state); 8 questions in one call returned 7.7–8.6× faster than 8 serial calls; batched and single answers differed no more than identical repeats (Noul 0.005, Score 0.010, Choice top never changed) blowxian, jev-fanout-bench via OpenRouter (provider TypeSafe, jev-1.13-20260917), 2026-09-23, $0.156; every request, response and checksum published Price and free output verified (Models, aliases, pricing, rate limits, context). 261 fits the docs' one-sentence Noul example (296 input_tokens) and Archer Hume's 268: a calculator counting only state + questions undercounts small requests. Serial timing; synthetic English tickets. Pattern: Speculative fan-out
Per-request formula (R² ≥ 0.999): 260 tokens, plus per question ~8 framing, ~1 per instruction word, ~8 per Choice option (21 with a six-word description), 8 per Score level. Characters per token: English 4.92, Spanish 4.44, Russian 1.78, Hindi 1.72, Arabic 1.49, Korean 1.44, Japanese 1.01, Chinese 1.00; JSON state ~2.35. Limits by bisection: state + longest question passed at 32,688, refused at ~32,813; the total passed at 65,388, refused at ~65,644. 1,000 questions in one call succeeded; 16–120 unrelated extra questions moved four target answers no more than repeats (median 426 ms at 4 questions, 436 ms at 124) same, round 2 479 billed requests + 8 over-limit probes; answers withheld, citing the MCA's competing-models term Limits verified: the brackets contain 32,768 and 65,536 (State: what you send Jev); "overhead counts toward neither" unverified. No question cap is documented (Archer Hume sent 5,000). His estimator at jevpricing.com claims 1.4% median error over 365 requests (unverified)
Translating the ticket moved yes/no answers 3–5× the English repeat noise; translating the questions too moved them further and cost 9–48% more tokens. Reordering Choice options never changed the top option; removing descriptions or paraphrasing moved probabilities 2–10 points. Over 960 reworded or translated answers, the lowest-confidence third of Choice answers moved 15× more than the top third (Spearman ρ −0.79) and held all five top-option changes (Score 7.5×, ρ −0.51) same same run English-first verified (Models, aliases, pricing, rate limits, context): keep questions in English. Order: Archer Hume (above) saw reversal move the top probability 0.84–0.89 → 0.93–0.96 and position change a reference option's hit rate. Stable winner, moving numbers: not a contradiction, but both unverified and the docs are silent. Re-check thresholds after any rewording
Latency flat to 100 questions: 3 / 25 / 100 / 400 questions → 521 / 1,242 / 4,017 / 15,417 input tokens and 712 / 690 / 714 / 1,336 ms (cold connection, Netherlands, 2026-09-17); warm 250–580 ms over 16 calls Colin de Vries, jevclient README table, one location (15,417 − 521) ÷ 397 ≈ 37.5 tokens per added question (our arithmetic), fitting the formula above and Archer Hume's flat-to-~100 curve
Server time p50 75–90 ms (x-envoy-upstream-service-time); warm round trip ~290 ms p50 from a machine 185 ms (TCP) away, ~1.2 s on a new connection. 16 requests at a time was fine; all 100 at once pushed round-trip p50 to ~2 s (server time still ~90 ms) and drew 529 Overloaded retries teyhouse, jev-secret-detection (no licence) 100-case set, jev-1.13.0 verified in direction: docs say "about 100 ms" (How to build software with System One); 529 is retryable (HTTP status codes, rate limits, retry semantics), so cap client concurrency. His "AWS us-west-2" is unverified. Accuracy: Failure reports: where Jev broke, lost, or was the wrong tool
19,772 requests in ~252 s at concurrency 16, 32.06M input tokens, ~$1.35 at list price bitnovus, jev-spam-eval full paired run, no API errors ~4,700 requests/min, ~127k tok/s (our arithmetic): above the documented 1,200 rpm, under 250k tok/s. jev-sec-bench's 662 in 22.7 s is ~1,750/min. Limits "can change without notice" (Models, aliases, pricing, rate limits, context); don't plan on headroom
Jev billed more input tokens than GPT-5.6 Terra for the same prompt: 360 vs 153 (8-way), 952 vs 828 (77-way), so per decision it was 40–49× cheaper than Terra, below the price-list ratio. Median 0.33 s; slowest of 791 calls 1.42 s Adel Dahani, AY Automate via OpenRouter, 3,955 calls, $1.53 total; agency blog Compare billed cost per decision, not price per token. Accuracy: Head-to-head: Jev against other models and methods

How to measure and what not to assume

Sources

Links are inline in each row. Raw captures (raw/x-repos/, raw/nate/, raw/community/) are in the private repo; Nate B. Jones's material is used with his permission and credited.