jev pricing and model limits

How Direct Jev Cost Is Calculated

TypeSafe charges the direct API by input tokens. At the published rate, one million input tokens costs $0.042. A workload sending 500 input tokens per request for one million requests uses 500 million input tokens, for a model-input cost of $21. This is arithmetic from the published price, not a quote for a complete production system.

The real cost also includes retries, provider markup if any, network and application infrastructure, labeling, monitoring and any fallback model or human review. Multiple questions share state in one request, which can reduce repeated state tokens, but questions themselves still consume input. Measure the serialized request rather than estimating from visible prose.

Worked Direct-API Input-Cost Examples

These calculations multiply total input tokens by the direct price verified on September 22, 2026. They exclude provider markup, retries, storage, orchestration and human review. Recheck the live model page before using them in a budget.

Context and Throughput Limits

The documented context budget is 64k tokens across state and all questions, with a separate 32k limit for state plus the single longest question. Hitting the maximum is rarely a quality target: TypeSafe warns that irrelevant long state lowers Jev 1.13 accuracy. Filter and structure state around the decision.

The published direct limits are 250,000 tokens per second and 1,200 requests per minute. TypeSafe explicitly says these are adjusting dynamically while demand and capacity change. Treat them as a dated snapshot and handle 429 responses using the SDK retry behavior or documented API guidance.

Shared State Changes the Cost of Multiple Questions

TypeSafe documents that multiple questions can share one state in a request and are evaluated independently in parallel. If five questions each need the same 800-token state, one combined request can avoid serializing that state five times. The questions themselves and request envelope still contribute input tokens.

Batch only questions that truly use the same evidence. Adding unrelated questions can enlarge the request, complicate attribution and tempt teams to send an overbroad state. The speculative fan-out guide covers parallel questions, limits and result handling.

Stable, Preview and Pinned Model IDs

Aliases can move without an application code change. Use an alias while experimenting, but pin the version behind a calibrated evaluator or consequential policy. The response reports the model that actually answered; log it and replay the evaluation before upgrading.

Direct API vs Gateway Pricing

Cloudflare publishes a Jev listing. An OpenRouter Jev route page exists, but its public models API did not list TypeSafe/Jev when checked September 22, 2026. Availability, request format, regions, quotas and price can differ from TypeSafe’s direct API. Choose a provider only after verifying and testing the complete path.

Model Price Is Only One Line of the Workflow Budget

A model with a low token price can still be expensive if poor thresholds send most traffic to review. Compare expected cost at the same quality and automation coverage. The human-review comparison provides a capacity calculation.

Tokens per Second and Requests per Minute Constrain Different Workloads

At 1,200 requests per minute, evenly distributed traffic averages 20 requests per second. At 250,000 input tokens per second, a 1,000-token request would reach the token limit near 250 requests per second before considering bursts. In this simplified example, the request limit binds first. Larger requests can make token throughput bind first.

Real traffic is bursty, and the published limits are dynamic. Load-test within authorized conditions, cap concurrency, honor rate-limit responses and queue asynchronous evaluation when immediate results are unnecessary. The retries and rate-limits guide covers the application control loop.