TypeSafe Launches Jev, a $0.042-per-Million-Token AI Model That Makes Decisions Instead of Writing Text
TypeSafe AI emerged from stealth on September 15 with Jev, the first public model in a category the company calls “System One Models.” Jev does not write prose or code. It accepts text or structured text state and returns predefined choices, scores, Boolean-style probabilities, and confidence values that software can consume directly. The company opened early access alongside the launch. [1]
The release comes from a team led by Diogo Almeida, whose previous work included OpenAI’s instruction-following and RLHF research. Investor DCVC separately announced that it led a $40 million Series Seed round in TypeSafe, describing the company’s goal as building models optimized for automation rather than human-facing conversation. [5]
A decision model instead of another chatbot
TypeSafe’s argument is that many production AI calls do not need generated language. Software often needs a bounded judgment: which queue should receive a ticket, how risky an action looks, whether a passage is relevant, or which tool an agent should use next.
Jev exposes three primitives. Choice selects an option from a declared set. Score evaluates an ordered rubric. Noul returns the probability that a statement is true. Multiple questions can share one state and are evaluated independently in parallel. The output shape is fixed in advance, so an application does not have to parse prose into a schema. [1]
That property should not be confused with guaranteed correctness. A typed answer cannot introduce an undeclared Choice key, but it can still choose the wrong key. TypeSafe says Jev is trained for calibrated decisions, meaning probability should correspond to accuracy across groups of predictions. Developers still need labeled evaluations and conservative fallback rules for their own workloads. [1]
Price, limits, and access
| Item | Current Jev 1.13 detail |
|---|---|
| Input price | $0.042 per million tokens |
| Output price | Free, according to TypeSafe |
| Context | 64k tokens per request; 32k for state plus the longest question |
| Input | Text only: string, JSON object, or arrays of text |
| Direct rate limits | 250,000 tokens per second and 1,200 requests per minute |
| Access | TypeSafe early access, official SDKs, and Vercel AI Gateway |
TypeSafe’s current model page lists Jev 1.13 as jev-1.13.0 and maps the jev-latest alias to that release. The company warns that aliases can move and answers can change, so production systems that tune confidence thresholds should pin the evaluated model version. It also says rate limits may change while it expands capacity. [2]
Vercel added Jev to AI Gateway under the model identifier typesafe-ai/jev. Its AI SDK evaluation API represents Jev’s outputs as Choice, Score, and Boolean questions and can request Zero Data Retention and No Training through gateway options. [4]
What developers can use it for
The clearest applications are high-volume decisions with a closed answer space. TypeSafe lists classification, routing, scoring, extraction from candidate values, guardrails, verification, and real-time software among the target workloads. Vercel highlights agent-loop decisions: choose a tool or subagent, decide whether to continue or retry, estimate risk before an action, and verify another model’s output. [1] [4]
For programming systems, that makes Jev more relevant to the coding-agent harness than to code generation. A generative model can inspect a repository and propose a patch; Jev could classify the failure, choose the next specialist, filter retrieved context, or decide whether the state is ready for review. Ordinary code would still own permissions, thresholds, and side effects.
TypeSafe reports that Jev was up to 193.6 times faster and 444.6 times cheaper than LLMs in its workflow evaluations. Those figures are vendor-run results, not independent benchmarks. The company says the comparisons probably represent the high end of real-world gains and acknowledges possible bias because members of its model-capabilities team created the workflows. [1]
The constraints are unusually important
Jev is deliberately narrow. It does not generate explanations, code, or open-ended answers. TypeSafe’s Jev 1.13 limitation page says the model can interpret instructions literally, struggles with counting and numerical precision, handles dates as text rather than ordered values, loses accuracy when state contains irrelevant detail, and can be influenced by adversarial content. It recommends keeping arithmetic and hard invariants in code and using a generative model when the task actually requires generation. [3]
The launch is therefore not evidence that decision models can replace general-purpose LLMs. It is a bet that software systems should stop paying a generative model to produce strings when the application only needs a bounded decision. Whether that bet works will depend on calibration against real data, the quality of each question and answer set, and the deterministic controls around the model.
For developers, Jev is worth watching because its listed price makes model calls plausible at a frequency that would be difficult to justify with a frontier reasoning model. The immediate opportunity is not a cheaper chatbot. It is a new component for the places where software has to make a fast judgment and code must remain in control.