Jev pricing & token economics.
per million input tokens
$0.00 output (free)
TypeSafe Jev is published at $0.042 per million input tokens ($0.000000042 per token, or $42 per billion tokens) with $0.00 (free) output tokens on TypeSafe direct, OpenRouter, and Cloudflare Workers AI.
Because Jev produces mathematical probability distributions (categorical Choice strings, continuous Score floats, or binary Noul probabilities) rather than autoregressive prose sequences, output tokens are zero-rated. API responses may report an output_tokens count in usage metadata even though Jev's published output-token price is $0/M. Billing is based entirely on the cumulative input token length of the shared state plus all question instructions and candidate criteria in questions.
There is no universal "per-request" price. Estimating per-request cost requires specifying the input token length of your state plus question schema.
This page covers Direct Jev (model identifiers typesafe/jev-1.13, jev-1.13.0, or alias ~typesafe/jev-latest), which is a structured System One decision model billed at $0.042/1M input tokens. If you are using OpenRouter's managed TypeSafe Jev Router (typesafe/jev-router), see the Jev Router pricing section: the router itself carries no separate model rate, but requests are billed according to whichever downstream model actually serves the completion.
Cost Calculator
Interactive workload & token estimatorEstimate your monthly or per-workload charges using either your projected request volume and average token size, or direct cumulative token volume. All calculations use the verified direct Jev rate of $0.042 per 1,000,000 input tokens ($0.00 output).
Calculated strictly from billable input tokens at the published rate of $0.042 per 1,000,000 tokens ($0.000000042/token). Output tokens are zero-rated ($0.00). Per-request cost varies directly with input token length.
Practical Workload Cost Examples
Exact arithmetic across common workload sizesBecause Jev bills strictly on input tokens, the table below translates $0.042/1M tokens into concrete, human-scale workload scenarios. Every per-request calculation explicitly exposes its assumed input-token count:
| Workload Profile | Assumed Input Tokens / Request | Monthly Requests | Total Billable Tokens | Monthly Cost | Effective Cost / Request |
|---|---|---|---|---|---|
| Fast Routing & Intent Guardrails | 250 tokens | 100,000 | 25,000,000 (25M) | $1.0500 | $0.0000105 |
| Standard Triage & Ticket Routing | 500 tokens | 10,000 | 5,000,000 (5M) | $0.2100 | $0.0000210 |
| Standard Triage & Ticket Routing | 500 tokens | 100,000 | 50,000,000 (50M) | $2.1000 | $0.0000210 |
| High-Volume Triage & Classification | 500 tokens | 1,000,000 (1M) | 500,000,000 (500M) | $21.0000 | $0.0000210 |
| Detailed Multi-Question Evaluation | 1,000 tokens | 100,000 | 100,000,000 (100M) | $4.2000 | $0.0000420 |
| Detailed Multi-Question Evaluation | 1,000 tokens | 1,000,000 (1M) | 1,000,000,000 (1B) | $42.0000 | $0.0000420 |
| Document & Context Chunk Intake | 5,000 tokens | 100,000 | 500,000,000 (500M) | $21.0000 | $0.0002100 |
| Enterprise Context Evaluation | 10,000 tokens | 100,000 | 1,000,000,000 (1B) | $42.0000 | $0.0004200 |
Because Jev does not bill per API invocation or per output token, quoting a single "request price" without stating input tokens is mathematically invalid. Evaluating 500 input tokens costs $0.000021 (over 47,000 requests per dollar), whereas evaluating 10,000 input tokens costs $0.00042 (2,380 requests per dollar)—a 20× difference. When forecasting budgets, measure the average token count of your state text plus question schema.
Verified Provider Pricing Matrix
Status: Verified October 2026Direct Jev inference is available across multiple providers. TypeSafe direct, OpenRouter, and Cloudflare Workers AI publish verified base token rates of $0.042/1M input and $0.00 output. Vercel AI Gateway operates on pass-through provider pricing, though public surfaces have displayed inconsistent formatting ($0.042/1M vs. rounded $0.04/M):
| Provider | Model ID | Input Price / 1M | Output Price / 1M | Minimum / Billing Policy | Primary Supported Route |
|---|---|---|---|---|---|
| TypeSafe Direct API | jev-1.13.0 | $0.0420 | $0.0000 (Free) | Pure usage-based inference billing; no monthly base fee or minimum spend | POST /v1/systemone |
| OpenRouter | typesafe/jev-1.13 | $0.0420 | $0.0000 (Free) | Standard OpenRouter credit balance; zero token markup | POST /api/alpha/decisions |
| Vercel AI Gateway | typesafe-ai/jev | Pass-through (~$0.042 / 1M)* | $0.0000 (Free) | Pass-through provider pricing ($0 token markup); monthly included credits apply | POST /v1/evaluate |
| Cloudflare Workers AI | typesafe/jev | $0.0420 | $0.0000 (Free) | Standard Workers AI platform usage and billing pass-through | @cf/typesafe/jev |
- OpenRouter Native Decisions API: OpenRouter's primary native route for Jev is
POST https://openrouter.ai/api/alpha/decisions. For teams using the official TypeSafe client library (@typesafe-ai/sdk), OpenRouter also provides the compatibility routePOST https://openrouter.ai/api/v1/systemone. See OpenRouter Integration Guide for complete code examples. - *Vercel AI Gateway Rate Presentation: Vercel applies pass-through provider pricing with zero token markup. However, public Vercel surfaces have displayed inconsistent rate formatting (ranging from $0.042/1M to a rounded $0.04/M, alongside promotional credits). Check your active Vercel dashboard for your team's exact billed precision. Optional enterprise add-ons (such as team allowlists at $0.10/1k requests or OpenTelemetry trace drains) are platform features, not token markups.
Direct Jev vs. Jev Router: Billing Disambiguation
Correcting common pricing misconceptionsIt is critical to distinguish between Direct Jev and Jev Router when evaluating costs:
Direct Jev is a standalone System One decision model. It takes a shared state and a structured map of questions, returning typed discrete evaluations (probabilities, scores, categories). Billing is deterministic: exactly $0.042 per million input tokens and $0.00 output tokens.
Jev Router is a managed routing proxy on OpenRouter's Chat Completions endpoint (POST /api/v1/chat/completions). It accepts standard chat messages and uses Jev internally to select the optimal downstream model (such as Claude, GPT-4o, or DeepSeek) and reasoning effort.
How Jev Router is billed: OpenRouter charges no platform routing fee ($0.00 markup for the routing layer itself). However, each request is billed at the standard published token rates of the downstream model that actually serves the completion. Jev Router has no fixed price of its own. Earlier catalog displays of "$0.00 / 1M prompt and completion tokens" reflected zero router platform surcharge, not free downstream text generation.
For implementation instructions and downstream model resolution patterns, read our dedicated TypeSafe Jev Router reference.
Context Limits & Batching Economics
Direct TypeSafe API vs. OpenRouter Context EnvelopesApplies to the cumulative sum of tokens in the shared state block plus all questions included in the request. If you batch 20 questions against a 20k-token state, the total request must fit within 64,000 tokens.
A separate hard ceiling applies to the shared state plus the single longest individual question. Regardless of how few questions are passed, this pair cannot exceed 32,000 tokens.
OpenRouter documents Jev 1.13 as having a flat 32,000-token context window in its model catalog. While TypeSafe exposes the 64k total request budget on direct endpoints, OpenRouter publishes a 32,000-token ceiling for unified compatibility.
Because Jev bills only input tokens, evaluating multiple questions against the same state in a single request amortizes the state token cost across all questions. Consider evaluating 10 classification questions against a 2,000-token document:
- Single Batched Call: 2,000 state tokens + (10 questions × 50 tokens) = 2,500 total input tokens = $0.000105.
- 10 Unbatched Calls: 10 × (2,000 state tokens + 50 question tokens) = 20,500 total input tokens = $0.000861.
In this scenario, unbatched requests cost 8.2× more for the same set of 10 questions because the 2,000-token shared state is re-sent and billed 10 separate times. While the exact multiplier varies based on your specific document length and question count, batching multiple questions against a shared state in a single request is the most impactful architectural pattern for minimizing token spend.
Jev vs. LLM Cost Economics
Interactive side-by-side token spend comparisonFor a classification or routing call, a generative LLM bills for the input prompt and every completion token it generates, plus reasoning tokens on newer reasoning models. In contrast, Jev bills only input tokens and outputs structured probabilities with zero output token charges. Enter your projected volume and your target LLM's published rates to compare monthly spend:
Token spend only, from the numbers you enter. It leaves out accuracy, latency, retries, and the extra format instructions an LLM prompt usually needs. Reasoning models bill their thinking as output tokens. Output presets are rough guides, not measurements.
Token spend only. Cheaper is not the same as better. Evaluate accuracy on your own labeled data before switching; see Jev vs. LLMs and benchmarks.
Pricing Caveats & Billing Rules
Practical details to prevent unexpected chargesAPI responses may report an output_tokens count in usage metadata even though Jev's published output-token price is $0/M across all verified providers. The final billed inference amount reflects exclusively input tokens.
Direct Jev has no per-seat licenses or model-level subscription tiers—token inference is billed strictly on consumption ($0.042/1M input). However, individual aggregator platforms may charge fees independent of model token rates. For example, OpenRouter applies payment processing fees when purchasing prepaid credits, and Vercel AI Gateway offers optional team and observability add-ons. Distinguish model token charges from account funding or platform infrastructure fees.
To ensure billing stability and predictable token counts, pin your requests to the specific release model string typesafe/jev-1.13 (or jev-1.13.0) rather than floating aliases, so that future model updates do not alter tokenizer efficiency unexpectedly.
Frequently Asked Questions
Direct answers to common pricing and billing queriesTypeSafe Jev costs $0.042 per million input tokens ($0.000042 per thousand tokens, or $42 per billion tokens). Output tokens are $0.00 / free on TypeSafe direct, OpenRouter, and Cloudflare Workers AI.
Yes. Because Jev is a non-autoregressive decision model that computes mathematical probabilities rather than generating variable text sequences, providers rate output tokens at $0.00. API responses may report an output_tokens count in usage metadata, but published output-token rates remain $0/M.
Cost depends strictly on the input token size of your request. For example, a 250-token routing check costs $0.0000105, a standard 500-token classification costs $0.0000210, and a detailed 1,000-token evaluation costs $0.0000420. One dollar buys approximately 23,800 requests at 1,000 tokens each, or 47,600 requests at 500 tokens each.
No. While OpenRouter adds no platform fee or routing markup for using typesafe/jev-router, each routed request is billed at the standard published token price of whichever downstream LLM (such as Claude, GPT-4o, or DeepSeek) is chosen to generate the completion text.
No. OpenRouter bills direct Jev (typesafe/jev-1.13) at exact parity with TypeSafe direct pricing: $0.042 per 1M input tokens and $0.00 per 1M output tokens, with zero token markup.
The most effective technique is multi-question batching. If you have multiple questions about the same piece of text, send them in a single API call rather than making separate calls. This ensures you only pay for the shared state tokens once rather than multiple times.
No. Both TypeSafe direct and OpenRouter provide usage-based model access without monthly seat licenses or subscription tiers for model inference. Note that payment processing fees may apply when funding prepaid credits on aggregators like OpenRouter, but there is no recurring model retainer.