Jev Pricing and Token Costs
Understand TypeSafe Jev pricing: $0.042 per million input tokens, unmetered output, context limits, and an interactive calculator to estimate workload costs.
TypeSafe Jev uses an asymmetric pricing model: $0.042 per million input tokens ($42 per billion tokens) with completely free, unmetered output tokens ($0.00). Pricing applies equally across direct API access (api.typesafe.ai) and OpenRouter (typesafe/jev-1.13).
Interactive Workload Calculator
Model monthly inference costs for your expected token payload and call volume:
Calculate Your Jev Workload Cost
Billed strictly at $0.042 per million input tokens ($42 per billion). Output tokens are unmetered ($0.00).
LLM comparisons assume prompt + 30 output tokens of structured JSON extraction. Equal token counts do not imply equal task quality; generative LLMs stream prose/JSON text while Jev evaluates structured decision heads directly.
* Estimate based on TypeSafe's published early-access rate of $0.042/M input tokens. Output pricing is $0.00 (unmetered output tokens). Excludes network transit costs.
Pricing Structure Breakdown
Unlike chat models where developers pay both input token fees and high completion token fees, Jev uses an asymmetric zero-output pricing model:
| Billing Metric | Rate | Notes |
|---|---|---|
| Input Tokens | $0.042 / 1M tokens | Billed on the combined length of the state payload and all question definitions ($42 per billion tokens). |
| Output Tokens | $0.00 (Free / unmetered) | Classifications, rubric scores, and probability values incur no generation billing. |
| Context Limit | 64,000 tokens | Total request budget (combined state payload plus longest question capped at 32,000 tokens). |
| Fine-Tuning | Not currently offered | Jev relies on prompt criteria and in-context state calibration. |
| Minimum Spend | None documented | Usage-based early-access billing via console.typesafe.ai. |
Cost Scenarios Across Common Workloads
To illustrate how the $0.042/M input rate translates into real operating costs, consider four common production scenarios:
1. Inbound Customer Support Triage
- Volume: 50,000 tickets/month
- Payload: 400 tokens (customer email + routing rubric)
- Total Tokens: 20,000,000 tokens
- Monthly Cost: $0.84
2. High-Frequency Real-Time Content Moderation
- Volume: 2,000,000 user comments/month
- Payload: 150 tokens (user submission + policy guidelines)
- Total Tokens: 300,000,000 tokens
- Monthly Cost: $12.60
3. Agent Tool Call Guardrails
- Volume: 500,000 tool executions/month
- Payload: 600 tokens (planned tool command + execution history + security criteria)
- Total Tokens: 300,000,000 tokens
- Monthly Cost: $12.60
4. Continuous Control Loop / Game Automation
- Volume: 10 requests/second continuous (25.9M requests/month)
- Payload: 250 tokens (game frame telemetry + valid action space)
- Total Tokens: 6.48 billion tokens
- Monthly Cost: $272.16
Measured Economics: 77-Choice Intent Routing Benchmark (September 2026)
In our September 2026 controlled benchmark on Banking77 intent classification (385 test samples across 77 intents), we tracked exact token consumption and billed costs for both TypeSafe Jev (jev-1.13.0) and Google Gemini 3.8 Flash (gemini-3.8-flash) executing the identical bounded classification task.
| Economic Metric | TypeSafe Jev (jev-1.13.0) | Google Gemini 3.8 Flash (gemini-3.8-flash) |
|---|---|---|
| Evaluated Decisions | 385 decisions | 385 decisions |
| Input / Prompt Tokens Billed | 622,374 tokens | 11,063 tokens |
| Visible Candidate Tokens | N/A (non-generative) | 2,044 tokens |
| Thinking Tokens | N/A | 42,664 tokens |
| Total Billed Output Tokens | 0 tokens (unmetered rate) | 44,708 tokens |
| Average Input Tokens per Call | ~1,617 tokens | ~29 tokens |
| Average Billed Output Tokens per Call | 0 tokens (unmetered rate; ~754 reported) | ~116 tokens (~5 candidate + ~111 thinking) |
| Official Rate Card | $0.042 / 1M input, $0.00 output | $0.75 / 1M prompt, $3.75 / 1M output (including thinking) |
| Total Measured Inference Cost | $0.0261 USD | $0.1760 USD |
| Measured Cost per 1,000 Decisions | $0.0679 | $0.4570 |
Why Choice Questions Consume Input Tokens for Candidate Criteria
In Jev's Choice primitive, each question includes a criteria dictionary describing the valid options. Because Jev tokenizes the prompt instructions, criteria keys, and descriptions as input context, asking Jev to choose between 77 separate intents required ~1,617 input tokens per evaluation.
In contrast, Gemini's structured output schema (responseSchema) with 77 enum strings was reported by Google's API at an average of ~28.7 prompt tokens per query. Google's API charges for input tokens and candidate tokens, plus any thinking tokens generated during inference.
How Output Token Pricing Impacts Routing Economics
Although Jev reported an average of 754 output tokens per response in its usage metadata (representing full probability distributions across all 77 classes), TypeSafe billed exactly $0.00 for those output tokens.
Across this 77-candidate benchmark, TypeSafe Jev cost $0.0679 per 1,000 decisions versus $0.4570 for Gemini 3.8 Flash because Gemini generated thinking tokens billed at Google's $3.75/M output rate. Jev also delivered decisions at an observed median latency of 188ms compared to Gemini's 1,204ms. This cost ratio applies strictly to this 77-intent configuration; workloads with fewer candidate options, different schema sizes, or models without thinking tokens will produce different cost comparisons.
Download the raw benchmark dataset: /data/jev-banking77-benchmark.csv.
Architectural Cost Framing: Decision Evaluators vs. Generative Models
When evaluating whether to replace small generative models (such as Gemini Flash) with Jev for classification or routing, developers should consider architectural fit rather than pure token math:
- Task Scope: Jev is purely a decision evaluator. It cannot produce freeform text, explain its reasoning, or rewrite content.
- Token Equivalence: Identical token volumes do not imply identical task capability. Generative models utilize multi-token autoregressive decoding to synthesize answers, whereas Jev evaluates state against predefined question options in a single forward pass.
- Cost Efficiency: Where a workflow strictly requires categorical routing, rubric scoring, or binary gating, Jev eliminates both output token charges and JSON parsing overhead, operating within TypeSafe's reported 70ms to 500ms service range (measured 188ms p50 on Banking77) at $0.042 per million input tokens.
Account Access and Quotas
During the current early-access phase, accounts generated via console.typesafe.ai operate under dynamic quota envelopes:
- Request Concurrency: Up to 1,200 requests per minute early-access ceiling.
- Throughput Ceiling: Up to 250,000 tokens per second early-access ceiling.
- Access Channels: Direct API access via TypeSafe console waitlist and instant on-demand access through OpenRouter.
Related Context
TypeSafe Jev: Architecture, Primitives, and Operating Guidelines
entityJev vs. Structured Outputs: Controlled Banking77 Benchmark
comparisonHow to Choose an AI Model for a Software Task
conceptSources & Documentation
- TypeSafe AI Official Pricing Page — System One Token Rates and Billing Terms, retrieved Sept 2026.
- OpenRouter Model Card: typesafe/jev-1.13 — OpenRouter gateway specifications and token pricing.
- TypeSafe AI Documentation: System One Quotas — Official documentation covering early-access rate limit envelopes.