Evidence Prices as of October 12, 2026
The cost ladder
The same judgment costs wildly different amounts depending on who makes it. With Jev as 1×, here's what one decision costs as you climb from decision models to small LLMs, to frontier LLMs, to a person.
One decision, normalized
Bars use a logarithmic scale, from 0.5× to 10,000×, so equal lengths mean equal multiples.
How these numbers were built
The reference decision is a typical mail-rubric request: about 1,500 input tokens on Jev, several questions over one email. At $0.042 per million input tokens, that's $0.000063, our 1×.
| Option | Price per million tokens (in / out) | Basis | Per decision | Multiple |
|---|---|---|---|---|
| Jev 1.13 | $0.042 / free | Measured: ~$0.00006 per mail request, ~$0.0001 per 13-question wine request | $0.000063 | 1× |
OpenAI Decisions (gpt-6-luna) | $0.10 / free | Measured: 9,314 calls, 17.8M tokens, $1.78; tokens per request matched Jev's on the same wines | $0.00015 | 2.4× |
| Claude Haiku 5.5, effort low | $0.10 / $0.50 | Measured: ~$0.0004 per mail request (2.5k in), ~$0.00055 per wine request (3.5k in, ~400 out). Requests run ~2× Jev's tokens (framing plus a newer tokenizer), and output is billed | $0.0004 | 5–7× |
| Claude Sonnet 5.5 | $2 / $10 | Estimated: Haiku's request shape (2,500 in, 400 out) at Sonnet prices | $0.009 | ~140× |
| Claude Opus 5.5 | $4 / $20 | Estimated, same shape | $0.018 | ~290× |
| Claude Fable 5.1 | $10 / $50 | Estimated, same shape | $0.045 | ~710× |
| Open weights on your Mac | — | No per-call charge; electricity and 2–18 s per request | ≈$0 | ≈0× |
| Open 2B on a rented NVIDIA L4 | ~$0.80 per GPU-hour | Estimated: 0.32 s per request serially, ~11,000 requests an hour, only if kept busy | ≈$0.00007 | ≈1.1× |
| A person | — | Illustrative: 30 seconds at $60 an hour | $0.50 | ~7,900× |
Read the estimates as floors. The frontier rows assume a frontier model answers the same small question with the same small output. In practice the LLM tier does bigger work (a deep read, a plan audit, a fix), with longer prompts and more thinking at higher effort, so one LLM escalation often costs 10× the per-decision figure above. That's another reason the cascade matters: the LLM's time is for problems only it can solve.
Two other prices to know. Microsoft's hosted Decision-1 (launched October 9) is listed at $0.042 per million input tokens, the same as Jev; we haven't tested it. Cloudflare's open Clef models are hosted on Workers AI at $0.24 (Clef) and $0.09 (Clef-flash) per million input tokens. Batch APIs halve the Claude prices (Haiku 5.5 on Batch: $0.05 / $0.25) when you can wait.
What a thousand decisions cost
| Option | 1,000 decisions | 1 million decisions |
|---|---|---|
| Jev 1.13 | $0.06 | $63 |
| OpenAI Decisions API | $0.15 | $150 |
| Claude Haiku 5.5 (low) | $0.40 | $400 |
| Claude Sonnet 5.5 (est.) | $9 | $9,000 |
| Claude Opus 5.5 (est.) | $18 | $18,000 |
| Claude Fable 5.1 (est.) | $45 | $45,000 |
For scale: the operations agent's decision model cost under $0.10 per overnight session. A full day of wine-cookbook runs, including tests, came to $0.54 on Jev. The whole Haiku benchmark (15,039 calls) cost $7.57; the Decisions API benchmark (9,314 calls) cost $1.78.
Calculator: what does your cascade cost?
Set how many decisions you make a day and where they end up. Every decision that leaves code is asked of the decision model first; escalations also pay for the LLM, and the person tier pays for both plus the person.
Two things fall out quickly. Once people are in the loop, their time dominates every other cost, so the human tier should see only true preference and stakes. And the decision model's share barely moves the bill. What moves it is how many decisions escalate past it. That's why the push-down loop, which turns each escalation into a rule or a context line, is the real cost lever.
Latency climbs a similar ladder. Hook Chain's latency ladder normalizes to one pass-through Claude Code hook (about 42 µs): a call to Jev is about 14,000× that, and Haiku about 60,000×.
Choosing by cost alone is a mistake. Haiku costs 6× Jev but wins on reasoning-heavy judgments; the Decisions API costs 2.4× but is the fastest hosted option. See when to use which.