A coding agent is about to run db:reset. If the task was to reset the database, that command is correct. If the task was to add a column, it’s a disaster. A regex blocklist can’t tell the difference, and a second call to a frontier model adds seconds and cents to every tool call. That’s the gap TypeSafe’s Jev fills: a model that returns typed decisions instead of text, fast enough to sit inside a request path.
What Jev actually returns
Jev is a System One model from TypeSafe AI, founded by Diogo Almeida, a co-author of the InstructGPT paper. It emerged from stealth on September 15, 2026, with $40 million in seed funding. The model is named after William Stanley Jevons, whose paradox holds that cheaper resources get consumed more. The bet: decisions cheap enough to make ten times a second will get made ten times a second.
Every call sends a state (text or JSON) plus a set of questions. Each question is a Choice, a Score, or a Noul (a yes/no probability). All questions are answered in parallel in one call. The response is typed: for a Choice, you get the chosen option, probabilities for every option, and a confidence score. No text to parse, no JSON repair, no retry loops.
Pricing is $0.042 per million input tokens with free output, and TypeSafe reports 70 to 500 ms end-to-end latency. The speed comes from skipping autoregressive generation entirely—the model reads the state once and produces every answer in the same forward pass.
The hallucination claim is narrower than it sounds
TypeSafe’s chart puts Jev at 0% hallucination, but the launch post admits that number is not empirical: schema matching is guaranteed, so they wrote 0. The model cannot return an option outside your list. It can absolutely return the wrong option from your list. As the most-replied Hacker News comment put it, it can’t emit an invalid type, but it can still emit a wrong valid value. Treat “type-safe” as a property of the output format, not a promise about accuracy.
Where the speed and cost numbers hold up
TypeSafe’s homepage claims 193.6x faster and 444.6x cheaper than frontier LLMs, but those figures come from four workflow evals built by TypeSafe staff, with a reference biased toward OpenAI and Anthropic models. The launch post’s own caveats say the gains are “on the higher end” of what to expect.
The best independent numbers so far come from Every’s head of evals, Mike Taylor. He ran 37 documents through 21 questions each—777 judgments—in under 0.7 seconds for about a quarter of a cent. Every’s CEO Dan Shipper then gave Jev and Fable 5.1 the same four writing checks on twelve passages. Jev took a median 0.35 seconds per passage against 8.83 seconds for Fable, at roughly 580x lower cost. Jev caught six of seven planted defects; Fable caught all seven.
That’s the honest summary: much faster and much cheaper on classification-shaped tasks, slightly less accurate than a top reasoning model. The trade is worth it when you need thousands of judgments, not when a single wrong decision is expensive.
What developers shipped in the first 48 hours
The community moved fast. Within 36 hours, Vercel added Jev to AI Gateway with an evaluate method in AI SDK 7. Developers shipped a coding-agent guardrail called pi-warden, an MCP connector so Claude can call Jev directly, a Ruby integration, a DSPy fork that routes Signatures to Jev, and an open-weights clone of the interface using Qwen3.5-4B logits.
pi-warden is the clearest example of the use case that holds up. Before each bash, write, or edit, it sends the task, the agent’s stated plan, and the pending command to Jev with four typed questions: is this irreversible, is it off-task, does it mutate anything, and what scope is it. Jev answers all four in about 250 ms, and code decides whether to hold the call. Over 17,000 recorded calls it held 42 times, and roughly 88% of those holds were right.
Where Jev fits next to an LLM, not instead of one
The use cases that hold up are the ones next to an LLM: reranking, citation checks, judging tool calls, routing, and verification. Anything that needs generated text, code, or an explanation still belongs to a frontier model. Multi-hop reasoning, arithmetic, counting, and date math are also weak spots for Jev.
For builders, the practical question is whether you have a judgment your software needs to make fast, thousands of times, where the answer is a decision rather than a paragraph. If so, Jev’s calibrated confidence gives you a second axis to threshold in code—the answer says what, confidence says whether to act on it. That’s a different interface from a chat model, and it changes where you can afford to put judgment calls. If you’re already thinking about model selection for agent workflows, Jev adds a new option for the decision layer.
Sources
AI-assisted summary compiled from the sources above, reviewed by a human before publishing.