Decision model pricing calculator: what Jev, Clef, Decider, Luna and Drex Max cost for your workload
Enter how many evaluations you run a month, how long the text is and how many questions each call asks. The table estimates input tokens per call for the decision models we host, the monthly bill at each vendor’s list price, and the cost on JevStation. Token counts come from our own calls (2026-10-07 to 2026-11-20). List prices alone mislead: Liquid AI’s d1 models share a low price per token with Perplexity Decider but bill the text once per question.
JevTypeSafe · $0.042 / 100万入力
- 直接利用・月額
- $2.66
- JevStation・月額
- $137.10
- 1回の入力トークン
- 634
- 1回のクレジット
- 1
Clef-flashCloudflare · $0.09 / 100万入力
- 直接利用・月額
- $5.53
- JevStation・月額
- $137.10
- 1回の入力トークン
- 614
- 1回のクレジット
- 1
ClefCloudflare · $0.24 / 100万入力
- 直接利用・月額
- $14.74
- JevStation・月額
- $270.30
- 1回の入力トークン
- 614
- 1回のクレジット
- 2
DeciderPerplexity · $0.02 / 100万入力
- 直接利用・月額
- $0.83最安
- JevStation・月額
- $137.10
- 1回の入力トークン
- 413
- 1回のクレジット
- 1
Luna DecisionsOpenAI · $0.10 / 100万入力
- 直接利用・月額
- $6.84
- JevStation・月額
- $137.10
- 1回の入力トークン
- 684
- 1回のクレジット
- 1
Decision-1Microsoft · $0.042 / 100万入力
- 直接利用・月額
- $1.47
- JevStation・月額
- $137.10
- 1回の入力トークン
- 350
- 1回のクレジット
- 1
d1-ultraLiquid AI · $0.042 / 100万入力
- 直接利用・月額
- $3.66
- JevStation・月額
- $270.30
- 1回の入力トークン
- 872
- 1回のクレジット
- 2
d1Liquid AI · $0.02 / 100万入力
- 直接利用・月額
- $1.74
- JevStation・月額
- $137.10
- 1回の入力トークン
- 872
- 1回のクレジット
- 1
Drex MaxNace.AI · $0.06 / 100万入力
- 直接利用・月額
- $3.20
- JevStation・月額
- $137.10
- 1回の入力トークン
- 533
- 1回のクレジット
- 1
実測した呼び出しに当てはめたトークンモデルによる推定です。9モデルとも出力トークンは無料。直接利用の価格にゲートウェイ手数料は含みません。JevStation の列は、登録時の 15 クレジットを差し引いた後の買い切りクレジットパックの最安組み合わせです。
How the estimate works
Each model’s row uses a token formula fitted to nine or more real calls per model (three text lengths × one, three and five questions), then multiplies by the list price per million input tokens.
| Model | List price, input | Input tokens per call, our fit |
|---|---|---|
| Jev | $0.042 / M | 260 + 0.163 × characters + 43 × questions |
| Clef-flash | $0.09 / M | 29 + 0.168 × characters + 111 × questions ¹ |
| Clef | $0.24 / M | 29 + 0.168 × characters + 111 × questions ¹ |
| Perplexity Decider | $0.02 / M | 48 + 0.163 × characters + 40 × questions ⁴ |
| OpenAI Luna Decisions | $0.10 / M | 22 + 0.163 × characters + 139 × questions |
| Microsoft-Decision-1 | $0.042 / M | 0.163 × characters + 35 × questions ² |
| Liquid AI d1-ultra | $0.042 / M | questions × (37 + 0.169 × characters) ³ |
| Liquid AI d1 | $0.02 / M | questions × (37 + 0.169 × characters) ³ |
| Nace.AI Drex Max | $0.06 / M | 0.163 × characters + 96 × questions ⁵ |
¹ On Cloudflare Workers AI on 2026-10-04 the text part stopped counting at about 2,048 tokens because the rest was not read. When we retested on 2026-10-07, both on Workers AI and through OpenRouter, Clef read and billed the whole text, so the formula now applies without a cap.
² Fitted on 2026-10-10, the day after launch, on the same nine requests.
³ Fitted on 2026-10-16, the day after d1-ultra launched, on the same nine requests. d1 and d1-ultra counted exactly the same tokens on every request.
⁴ Fitted on 2026-10-25 for perplexity/pplx-decider-v1.2, on the same nine requests. The earlier v1, which we measured on 2026-10-07, read the text once per question (questions × (84 + 0.163 × characters)) and cost $0.04 / M; OpenRouter has since retired it.
⁵ Fitted on 2026-11-20 for nace-ai/drex-max, on the same nine requests (constant term −5, floored at 0).
The formulas are for English prose. Code, JSON and Chinese, Japanese or Korean text produce more tokens per character on every model.
Three workloads, priced
At list price, per 100,000 evaluations:
| Workload | Jev | Clef-flash | Clef | Decider | Luna Decisions | Decision-1 | d1-ultra | d1 |
|---|---|---|---|---|---|---|---|---|
| Router: 1 question, 300 characters | $1.48 | $1.71 | $4.56 | $0.27 | $2.10 | $0.35 | $0.37 | $0.18 |
| Ticket triage: 3 questions, 1,500 characters | $2.66 | $5.53 | $14.74 | $0.83 | $6.84 | $1.47 | $3.66 | $1.74 |
| Document review: 5 questions, 8,000 characters | $7.47 | $17.35 | $46.27 | $3.10 | $20.21 | $6.21 | $29.17 | $13.89 |
The cheapest model changes with the shape of the request. Liquid AI d1 is the cheapest router, with Decider second. Once there are several questions, Decider v1.2 is the cheapest by a wide margin: it has d1’s price per token but reads the text once. Microsoft-Decision-1 comes next, though it runs on an Azure endpoint with tight rate limits. d1-ultra costs about the same as Decision-1 for one question and climbs fastest after Clef, because it rereads the text per question. Accuracy was level across Jev, Decision-1, d1 and d1-ultra in our tests, so cost is not the only thing to check.
Sources
- List prices. TypeSafe, Cloudflare Workers AI and OpenRouter (vendor)
- Token formulas. Our calls on 2026-10-04 (Clef, Workers AI; cap retested 2026-10-07) and 2026-10-07 (Decider and Luna Decisions, OpenRouter) and 2026-10-10 (Microsoft-Decision-1, OpenRouter) and 2026-10-16 (Liquid AI d1 and d1-ultra, OpenRouter) and 2026-10-25 (Decider v1.2, OpenRouter) and 2026-11-20 (Drex Max, OpenRouter) (ours); Jev from What One Jev Evaluation Costs (ours)
- How these models compare on accuracy. Decision models compared and Jev vs Drex Max (ours)
よくある質問
- Which decision model is the cheapest?
- It depends on the number of questions. For one question per call, Liquid AI d1 ($0.02 per million input tokens) is the cheapest we have measured. From two questions up, Perplexity Decider v1.2 wins: the same $0.02 per million, but it reads the text once, while d1 and d1-ultra read it again for every question. Microsoft-Decision-1 is next ($0.042 per million, also reads the text once). The calculator shows the numbers for your workload.
- Why do token counts differ between models with similar prices?
- Tokenizers and prompt formats differ, and one model bills differently. We measured: Jev ≈ 260 + 0.163 × characters + 43 × questions; Clef and Clef-flash ≈ 29 + 0.168 × characters + 111 × questions, for the whole text; Luna Decisions ≈ 22 + 0.163 × characters + 139 × questions; Decider v1.2 ≈ 48 + 0.163 × characters + 40 × questions; d1 and d1-ultra ≈ questions × (37 + 0.169 × characters).
- Are output tokens charged?
- No. All nine decision models we list bill input tokens only; the answer, a few probabilities, is free.
- How does JevStation price these models?
- Per evaluation in credits, not per token: 1 credit for a standard evaluation (up to 8,000 characters and five questions) and 3 for a large one, doubled for Clef and d1-ultra. Credits come in one-time packs, and the first 50 are free on sign-up.
- Is going direct cheaper?
- At high volume and list price, usually yes: you pay only tokens. You also write and run the integration for each vendor, and Jev direct needs TypeSafe waitlist access. JevStation trades a per-call margin for one key, one request format, a playground, batch runs and refunds on failure.
関連
- 意思決定モデル比較Jev・Clef・Decider・Luna・Decision-1・Liquid d1・Drex Max:精度、キャリブレーション、コスト、レイテンシ。
- Jev vs DeciderPerplexity Decider against Jev: same contract, measured accuracy, calibration, latency and token counts.
- Decisions API alternativesOpenAI Decisions API without the waitlist, plus Jev, Clef and Perplexity Decider: measured accuracy, cost, and how to port a request.
パイプラインに組み込む
新規登録で 15 クレジットを無料進呈。独自の質問セットを保存し、同じ評価を API から呼び出せます。