What does Jev (TypeSafe AI) cost per decision? A worksheet against your current LLM
Jev is billed per input token and output is free, so the cost of one decision is set by how many input tokens a request is counted as. That count includes a fixed overhead that third parties measured at roughly 250–300 tokens, even for a one-line request. Measure your own count, enter it below, and compare with the model you use now.
Cost per decision = input tokens per request × $0.042 per million ÷ decisions per request. At 300 input tokens and one question per request that is $0.0000126 per decision, or $12.60 per million decisions (arithmetic example, not a measurement). Batching several questions about the same state into one request lowers the cost per decision, because the state is billed once per request. Documented TypeSafe models page and parallel-questions cookbook, checked 28 Sep 2026
Worksheet
How the worksheet calculates
| Output | Formula | Notes |
|---|---|---|
| Requests per month | decisions per month ÷ decisions per request | Batch questions about the same state to lower this. |
| Billed requests | requests × (1 + retry rate) | Assumes every retried request is billed in full. TypeSafe does not document whether failed requests are billed, so this is the worst case. |
| Jev cost | billed requests × input tokens per request × $0.042 ÷ 1,000,000 | Output tokens are ignored because they are free on TypeSafe direct. |
| Fallback cost | decisions × fallback share × your fallback cost per decision | You enter the fallback cost; this site publishes no competitor prices. |
| Break-even fallback share | 1 − (Jev cost per decision ÷ your current cost per decision) | Above this share, Jev plus fallback costs more than using your current model for everything. |
The tool runs only in your browser, sends nothing, and calls no model. It uses one price constant: $0.042 per million input tokens, from the TypeSafe models page checked on 28 September 2026 and rechecked, unchanged, on 7 October 2026 (05:19 UTC). Update your own figures when the price or your prompts change.
What TypeSafe bills
| Question | Answer | Label and source |
|---|---|---|
| What is billed? | "Charged per input token. Output tokens are free." $0.042 per million ($42 per billion). | Documented models page, 28 Sep |
Responses show output_tokens. Do I pay for them? | No, not on TypeSafe direct. The API examples report 18–34 output tokens, but the price line says output is free. Ignore output_tokens when costing the direct route. | Documented API reference, 28 Sep |
| Which field do I measure? | usage.input_tokens, "Number of input tokens used" (Python: or None when the API did not report it). On OpenRouter, every Jev response also carries usage.cost. | Documented SDK reference; OpenRouter Jev guide, 28 Sep |
| Is the state billed once or per question? | Once per request: "N single-question calls pay for it N times … the batched call pays once." | Documented parallel-questions cookbook, 28 Sep |
| Is question and criteria text billed? | Not stated. Question keys are "not sent to the underlying model". The cookbook's cost figures suggest about 57 tokens per extra question in one example, but that depends on an assumed price constant the pool could not see. | Unverified derived from the cookbook; not a documented rule |
| Is there a minimum per request? | Not documented. TypeSafe's own API examples report 296–318 input tokens for a one-sentence state and one short question. | Documented example values, not measurements |
What others measured
These are third parties' own measurements, not reruns by the pool. Reported
| Operator and date | Task, size, route | What they report | Published data |
|---|---|---|---|
| AY Automate, 20 Sep (run 19 Sep) | 791 labelled decisions in 3 tasks, via OpenRouter | Mean input tokens, Jev vs GPT-5.6 Terra on the same prompt: 360 vs 153 (8-way intent), 952 vs 828 (77-way intent), 324 vs 127 (prompt injection). | summary.json published; raw requests not inspected |
| Opper, 25 Sep | 362 items; identical request bodies sent to Jev and Kev | "Jev adds a fixed charge of about 257 input tokens to every request." A one-line message with one question: 280 tokens on Jev vs 23 on Kev. Opper says TypeSafe-direct counts matched. | Benchmark repository published. Opper sells a Jev route and has a commercial interest. |
| NavyaAI, 21 Sep (updated 25 Sep) | AG News, n = 200, TypeSafe direct | 442 input tokens per decision on average; $18.57 per million decisions, against $16.30 for their own DIY setup on a small LLM. | Public dataset; no Jev request code seen |
| madewithjev, 28 Sep | 15 builds that published both volume and total cost | Median $0.000068 per decision. | Builder-reported figures; the site says they are not audited |
Arithmetic check: 442 tokens × $0.042 per million = $18.56 per million decisions, which matches NavyaAI's figure. The pool publishes no competitor model prices; enter your own in the worksheet.
How to lower cost per decision
- Measure
usage.input_tokenson 50–100 real requests; do not assume the size of your prompt text. - Ask all the questions about one state in one request, so the state and the fixed overhead are billed once.
- Trim the state to what the decision needs; this also helps accuracy (see limitations).
- Keep retries bounded (see errors and rate limits) and track the fallback share, which usually dominates total cost.
- Buy from TypeSafe or a named gateway; lookalike resellers charged 3–16× the list price (see official vs reseller).
What was not verified
- The pool made no Jev call and saw no bill; the worksheet uses the list price only.
- Whether question and criteria text, and failed or retried requests, are billed is not documented.
- How OpenRouter's Jev Router bills the model it routes to is not documented on the pages checked.
- All token counts and per-decision costs from third parties are their own measurements on their own tasks.