{"slug": "typesafe-ai", "title": "Typesafe AI", "summary": "TypeSafe AI announced Jev, a new \"System One Model\" built on a new architecture, sampler, and training algorithm it calls Reinforcement Learning for Calibrated Decisions (RLCD), which returns typed decisions with calibrated confidence probabilities instead of chat text. TypeSafe AI claims Jev is 193.6x faster and 444.6x cheaper than LLMs on System One workflows, completing a task in 0.114s for $0.000081 versus 8.566s and $0.013880 for LLMs, and prices Jev at $42 per billion input tokens, which it says is 238x lower than Claude Fable 5.1's input price. TypeSafe AI says the model targets automation by letting software act autonomously when confidence is high and escalate for review when it is not.", "body_md": "∵ ⩆\n\n⩆ ∵\n\n# We took the opposite research direction\n\nnot chat\n\nReinforcement Learning from Human Feedback (RLHF) has led to LLMs that are optimized for human preferences. This has led to models that are superhuman at instruction following, and are what we now call “chat.” Yet RLHF creates inherent issues such as mode dropping, overconfidence, and lack of reliability. These flaws mean that LLMs require humans-in-the-loop.\n\na new model\n\nWe built a new class of models, System One Models, to be natively used by machines. We’re building with a new architecture, a new sampler, and a new training algorithm: Reinforcement Learning for Calibrated Decisions (RLCD).\n\nDecisions, not strings\n\nTyped outputs that software can act on.\n\ncalibrated confidence\n\nEvery decision includes an estimate of how confident the model is.\n\nmore like code\n\nReliable, fast, and type-safe.\n\n[b.64]\n\nZmxvYXQgUV9yc3FydCggZmxvYXQgbnVtYmVyICkKewoJbG9uZyBpOwoJZmxvYXQgeDIsIHk7Cgljb25zdCBmbG9hdCB0aHJlZWhhbGZzID0gMS41RjsKCXgyID0gbnVtYmVyICogMC41RjsKCXkgID0gbnVtYmVyOwoJaSAgPSAqICggbG9uZyAqICkgJnk7CglpICA9IDB4NWYzNzU5ZGYgLSAoIGkgPj4gMSApOwoJeSAgPSAqICggZmxvYXQgKiApICZpOwoJeSAgPSB5ICogKCB0aHJlZWhhbGZzIC0gKCB4MiAqIHkgKiB5ICkgKTsKLy8JeSAgPSB5ICogKCB0aHJlZWhhbGZzIC0gKCB4MiAqIHkgKiB5ICkgKTsKCXJldHVybiB5Owp9\n\n# 193.6x Faster,\n\n444.6x Cheaper.\n\n*based on workflows for System One tasks [(proof)](./blog/introducing-system-one-models-and-jev)\n\nTypeSafe AI\n\nCost $0.000081\n\nCompleted in 0.114s\n\nLLMs\n\nCost $0.013880\n\nCompleted in 8.566s\n\nWatch the real video\n\nBuilt for automation\n\nJev returns typed decisions with calibrated probabilities, so your software can account for uncertainty. Set the thresholds for when it acts autonomously and when it asks for review. Combine those decisions in code to build larger workflows, with control over how the intelligence is used.\n\n## Jev’s intelligence per dollar is literally off the charts.\n\nMachine-Native Intelligence\n\nLLMs produce words for people. Jev produces typed decisions and is more like code: reliable, fast, self-consistent, and type-safe.\n\nZero Hallucinations\n\nEvery Jev decision comes with a confidence estimate, so your software can act when confidence is high and escalate when it is not.\n\nJev.Cost\n\n## $42\n\nPer Billion input tokens.\n\n## 238x\n\nLower input price than Claude Fable 5.1\n\n## Come Build With Us\n\n∵ ⩆\n\n⩆ ∵\n\nTypeSafeAI Blog\n\n⩆ ∵\n\nThoughts\n\nThe Bitterest Lesson\n\nTL;DR: Compute drives progress in AI, but what good is progress if you are not doing the right task!\n\nRead More\n\nThoughts\n\nAI: too good to be true, too bad to be useful | TypeSafe AI\n\nRLHF-trained language models please humans and assist rather than make reliable autonomous decisions. What comes next?\n\nRead More\n\n∵ ⩆\n\n⩆ ∵\n\n∵ ⩆\n\n⩆ ∵\n\n## We give a FAQ\n\nWhat are System One Models? What is Jev?\n\nSystem One Models are a new class of AI model built for decisions inside software. Jev is TypeSafe’s first public System One Model, optimized for automation. Send Jev structured questions and get typed decisions with probabilities and confidence that your software can act on.\n\nIs Jev just a smaller LLM?\n\nHow is this different from JSON mode or structured outputs?\n\nHow can Jev be so fast and inexpensive?\n\nAre these prices temporary or subsidized?\n\nWhat is Jev good at? Where does it struggle?\n\nCan Jev still get things wrong?\n\nIs Jev deterministic?\n\nHow do I get started or ask a question?\n\n[b.64]\n\niVBORw0KGgoAAAANSUhEUgAAABwAAAAcCAAAAABXZoBIAAAAl0lEQVR42mNgGPyg5u9/e1xyCV9+/7XDJVn/G7eky5vfl+U5sMvZPPn9Ow6XobP//t2LS07k7+/XTjjkFM7+/V2HS2PG7787+HHIBXz4fVAcl6F///6dj8vQ6b9//1bHIWdw9/fvNbg0vvr9+wgPLsm/v39H4pKb///vX3lcNj75+70HR4AzOPz+ewdn/OOVlDiIR5K6AACSCULwD4UI6QAAAABJRU5ErkJggg==\n\n∵ ⩆\n\n⩆ ∵\n\nVHlwZVNhZmUgQUkgSW50ZWxsaWdlbmNlIE5vdw==", "url": "https://wpnews.pro/news/typesafe-ai", "canonical_source": "https://typesafe.ai/", "published_at": "2026-09-15 21:03:13+00:00", "updated_at": "2026-09-15 21:39:25.394842+00:00", "lang": "en", "topics": ["artificial-intelligence", "ai-research", "ai-products", "ai-agents", "ai-startups"], "entities": ["TypeSafe AI", "Jev", "System One Models", "Reinforcement Learning for Calibrated Decisions", "RLCD", "RLHF", "Claude Fable 5.1"], "alternates": {"html": "https://wpnews.pro/news/typesafe-ai", "markdown": "https://wpnews.pro/news/typesafe-ai.md", "text": "https://wpnews.pro/news/typesafe-ai.txt", "jsonld": "https://wpnews.pro/news/typesafe-ai.jsonld"}}