TypeSafe's Jev returns typed decisions, not prose, with free output
A new 'System One' class trades prose for calibrated decisions at $0.042/M in — treat its benchmarks skeptically. Also: OpenAI begins testing in-chat ads.
Copy markdown
A model that answers with a decision, not a paragraph
TypeSafe's Jev, out Sept 15, is the first of what it calls 'System One' models: unstructured state in, one typed decision out. Three primitives cover most of it — Choice (pick an option), Score (rate on a scale), and Noul (a calibrated true/false) — each answer carrying a probability you can threshold on. It targets routing, guardrails, ranking, and extraction: the work that ends in a JSON object anyway.
Output tokens are free; input is $0.042 per million
Because the outputs are tiny typed values, TypeSafe meters only input — $0.042/M, output free — and clocks end-to-end latency at 70-500ms versus seconds for a frontier LLM doing the same classification. For high-volume routing or moderation where you fire a model on every event, that flips the cost curve: fractions of a cent per call instead of cents.
You can wire it in today — REST, SDKs, and a Claude Code skill
Access is REST (POST /v1/systemone) plus official Python and JavaScript SDKs, and there's already an agent skill for Claude Code so your coding agent can call Jev directly. It's early access behind a waitlist with a playground at console.typesafe.ai; grab a key and you can swap a pile of brittle 'ask GPT to return JSON' prompts for typed calls with confidence scores.
Read the 4,000x-cheaper claims with a raised eyebrow
Every headline multiplier — '193x faster,' '~4,000x less per call,' 76% accuracy within three points of GPT-6 and Opus 5 — comes from TypeSafe's own workflow evals, which it designed, and it concedes the pricing 'may be subsidized.' The primitive is genuinely new and worth testing on your own data; the leaderboard framing is marketing until someone independent reproduces it.
Elsewhere: OpenAI starts testing ads inside ChatGPT
OpenAI began piloting 'Sponsored Agents' — brand-backed agents that surface inside ChatGPT — alongside ad tooling with HubSpot and Shopify integrations, for select US advertisers. Nothing to build against yet, but it's the first concrete sign the free assistant you and your users lean on is becoming an ad surface; watch how it bleeds into the API and agent experiences.