Jev pricing, free plan and limits

Jev swaps open-ended text generation for typed decisions (a yes/no probability, a choice, or a score) that software can act on directly. A 64,000-token context, text-only input and per-account throughput caps suit classification, routing and scoring pipelines rather than chat, and it still has no public SWE-bench, GPQA or MMLU results.

Jev, TypeSafe AI's first System One Model, has no entry in the Amazon Bedrock, Vertex AI or Azure model catalogs as of October 7, 2026. Developers call it through TypeSafe's own API, open to everyone since September 27, or through OpenRouter, and it returns typed decisions with confidence scores in 70 to 500 milliseconds.

Provider: TypeSafe AI · Family: System One Models

More about TypeSafe AI on HokAI

Context window: 64,000 tokens

Input modalities: text · Output: text

About Jev

Jev is TypeSafe AI's first public model and the debut of what the company calls a System One Model: a new class of frontier model built to return fast, structured decisions rather than generated text. TypeSafe announced Jev on September 15, 2026, in a launch post from founder Diogo Almeida, who said the model runs on a new model architecture, a parallel sampler for efficiency, and a training method the company calls Reinforcement Learning for Calibrated Decisions (RLCD). Access started as a waitlist; on September 27, 2026 TypeSafe dropped it and opened sign-up to everyone. TypeSafe has not disclosed Jev's parameter count or further architectural detail. The hosted API is versioned; TypeSafe's documentation lists jev-1.13.0 as the current release, with the jev-latest and jev-preview aliases both pointing to it.

Unlike a conventional LLM, Jev does not generate text token by token. A single call sends a state (unstructured text or a structured JSON object describing a program's current situation) plus one or more typed questions, and Jev returns every answer from one parallel pass instead of sampling sequentially. TypeSafe documents three question kinds: noul (a yes/no question returning the probability of yes), choice (pick one option from a declared set, capped at 255 options, returning probabilities and a confidence score), and score (rate the state against an ordered list of descriptive levels, returning a continuous value with probabilities and confidence). Because the space of possible outputs is declared by the developer up front, TypeSafe says the model cannot return a value outside that schema, a guarantee about output structure rather than about the correctness of the underlying judgment.

Jev accepts up to 64,000 tokens per request, with the state portion plus the longest single question capped at 32,000 tokens, per TypeSafe's published API documentation. Input is text only: a string, a JSON object, or an array of text values. TypeSafe's own demos note that non-text data, such as a game screen, has to be pre-processed into a text or structured description first, since Jev does not accept images, audio, or video directly. English is its primary training language; the documentation says other languages, CJK scripts included, are handled less accurately and should be tested before production use. Throughput caps are per account and, in TypeSafe's words, adjusting dynamically while it brings on more GPU capacity, so treat the figures in the FAQ below as today's ceiling.

TypeSafe has not published Jev results on standard public benchmarks such as SWE-bench, GPQA, or MMLU, saying the model was not built for the open-ended reasoning or writing tasks those measure. Instead the company built its own workflow evals: a small set of decision workflows scored against a reference answer it computes as the average of two other frontier models. TypeSafe's launch post states that its widely repeated home-page claims of being 193.6 times faster and 444.6 times cheaper on these tasks come directly from that evaluation, while cautioning in the same post that the workflows were designed by its own capabilities team, that the reference answers are biased toward OpenAI and Anthropic's models, and that it cannot yet prove its pricing is not subsidized. Separately, TypeSafe states Jev's end-to-end response time as 70 to 500 milliseconds, against a stated 3 to 329 seconds for the frontier LLMs it compares itself against on similar tasks.

Jev bills only for input tokens, with output free of charge, which the company calls too cheap to meter; exact figures are in the pricing FAQ below. TypeSafe says this undercuts typical frontier-LLM input pricing by roughly an order of magnitude, and that those LLMs typically charge several times more for output than for input, a cost shape Jev does not share since it has no output charge at all.

On the cloud question, Jev sits outside the big model marketplaces. TypeSafe hosts it itself, and its trust center names Amazon Web Services as the cloud provider behind live requests, with Modal for inference compute and, from October 3, 2026, CoreWeave and Nebius as added subprocessors. Running on AWS hardware is not the same as a Bedrock catalog entry: there is no Bedrock model ID, so an AWS team wires Jev in as an outside HTTPS call from its own service, such as a Lambda function, the way it would any third-party API, unlike a catalog offering such as Claude in Amazon Bedrock. The second route is OpenRouter, whose OpenAI-compatible gateway carries Jev 1.13 under the ID typesafe/jev-1.13-20260917. Vercel has also published a community guide for calling Jev from its AI SDK. Teams that need everything inside one Google Vertex AI or Microsoft Azure contract will find models such as Claude Opus 5 there and Jev not.

TypeSafe positions Jev for what it calls AI-powered workflows: classification, routing, scoring, extraction, and branching logic embedded inside existing software, where a conventional LLM's free-form text output would need parsing and validation and could still land outside the expected shape. Its own public demos, a Doom-playing bot driven by a text description of game state and a Wikipedia-link-following Wikiracing bot, are built to show the model choosing among many discrete options quickly and cheaply rather than holding a conversation. TypeSafe explicitly does not position Jev for human-facing chat, open-ended writing, or general coding, jobs it says still belong to conventional LLMs with a person supervising the output.

Who should look elsewhere: if the job needs generated prose, image or audio understanding, a context window far past 64,000 tokens, or procurement through an existing cloud marketplace agreement, a conventional model from the models directory is the simpler choice. Gemini 3.5 Flash-Lite runs on Vertex AI with a context near one million tokens, and GPT-5.6 Luna is available on Azure; both write text, and both come from vendors with longer public track records. A team classifying a few hundred items a day may also find Jev's speed and cost advantage too small to justify adding a new supplier and a new integration path.

On governance, TypeSafe's documentation rules out training on customer traffic (details in the FAQ below), and its trust center lists a SOC 2 Type II report dated 2026. The company has still not published a system card, a training data cutoff date, or ISO 27001 or HIPAA coverage for the product.

Pricing

Input tokens: $0.042 per million ($42 per billion), as TypeSafe's docs listed them on the HokAI check of this page. Output tokens: free. TypeSafe's homepage puts that input rate 238 times below [Claude Fable 5.1](/hub/models/claude-fable-5.1)'s. No free tier: anyone can open an account at console.typesafe.ai and is billed per input token. Custom and enterprise plans add higher rate limits and zero data retention through sales, with no published price. [OpenRouter](/hub/tools/openrouter) charges the same $0.042 input and $0 output rates.

What a real job costs

Budgets: 20-page PDF = 30k in / 1k out · Support reply = 2k in / 300 out · Coding agent run = 200k in / 20k out. Computed from the vendor's per-token prices at render time; cached-input discounts are not applied.

Key Features

• Typed decision primitives: Three typed question kinds (noul, choice, score) let a single call return a yes/no probability, a pick from up to 255 declared options, or a continuous rating, always inside a developer-declared schema.
• Parallel, non-autoregressive sampling: Jev returns every requested answer from one parallel pass instead of generating tokens one at a time, which TypeSafe credits for its sub-second, real-time response speed.
• Calibrated confidence on every answer: Every choice and score response carries a confidence value between 0 and 1 that TypeSafe derives from the shape of the output probability distribution, meant to flag when a decision needs human review.
• Input-only pricing: TypeSafe charges only for input tokens and gives output away free, a cost shape unlike standard per-token chat model pricing (see the pricing FAQ for exact rates).
• Two access routes: Besides TypeSafe's own endpoint and its official Python and JavaScript SDKs, Jev 1.13 has been callable through OpenRouter's OpenAI-compatible API since September 18, 2026.

Pros

• TypeSafe reports sub-second, real-time response times fast enough for the live game and robotics demos the company has shown publicly.
• Charges nothing at all for output tokens, unlike standard per-token chat models that typically price output well above input.
• Output is constrained to a developer-declared schema (noul, choice, or score), which TypeSafe says makes a schema-violating response mathematically impossible.

Cons

• Accepts text only; images, audio, and video must be converted to a text or JSON description first, so it cannot process raw screenshots or audio directly.
• TypeSafe has not released independent SWE-bench, GPQA, or MMLU numbers, so how Jev's judgment holds up beyond the company's own workflow evals is still an open question.
• Not offered inside AWS Bedrock, Google Vertex AI or Azure, so teams standardised on one of those marketplaces must add TypeSafe or OpenRouter as a separate vendor, and there is no self-hosted option.

Visit Jev Official Page