Jev

Sources#

• Building Prod with Jev and LangGraph
• Inside the JEV Ecosystem: 13 Answer Verifiers on One Test Set
• Introducing System One Models & Jev

What it is#

The first public model from TypeSafe AI, launched in early access on 2026-09-15 by founder Diogo Almeida (formerly OpenAI, where he describes working on the instruction-following research behind ChatGPT) after two years in stealth. TypeSafe calls the class System One Models — after Kahneman's fast/slow distinction — and names Jev after William Stanley Jevons, betting that each order-of-magnitude drop in the cost of intelligence unlocks more than an order of magnitude more use. The pitch in one line: "a frontier-intelligence function call: unstructured state in, typed probabilistic decisions out." Everything below is from the launch post and is vendor-claim unless a second source is named.

• Not an LLM, and it gives up strings. The output space is declared in advance (bool, score, or choice per question); all outputs are produced in parallel in one query rather than token by token. Choices support a cardinality of up to 255; beyond that the vendor runs a two-stage score-then-choose, with a visible slowdown.
• Trained with RLCD — "Reinforcement Learning for Calibrated Decisions": the objective is epistemically honest probabilities on System One tasks, contrasted explicitly with RLHF (text raters prefer) and RLVR (programmatically checkable outputs, which the vendor says produce "spikey / non-robust intelligence" on judgment tasks). No algorithmic detail is published; training data is all synthesized in-house ("primarily a data research lab").
• Price and speed. $0.042 per million input tokens, output unmetered; 70–500 ms end to end (vs. 3–329 s quoted for frontier LLMs). The vendor concedes it cannot prove the price is unsubsidized.
• Type errors. 0% by construction — schema matching is guaranteed, so the 0% bar in its error-rate chart is stated as not empirical.

What independent measurement says#

The only third-party numbers are on Typed Decision Verifiers: on Proto_AGI's shared 2,018-item answer-correctness set (empirical), Jev scores AUC 0.7350, statistically tied with VIDRAFT's ZTC 397B for first and above GPT-5.2 (0.7148) at ~$0.024 vs ~$0.55 per 1,000 calls — but as a 20%-budget retry gate it moves end-to-end agent accuracy by −0.07pp, "repairing about as much as it damaged," because 216 of the 403 items it routed were already correct. The same study puts Convai's Laya (~0.015 s/call) at "roughly 140× faster than the hosted API we measured" — in context almost certainly Jev's, which would put it near ~2 s per call, well outside the vendor's 70–500 ms band (the vendor's own caveat is that its latency evals run from West-Coast laptops near the service; the study does not state its client location or input lengths). That study's independence is also less clean than first recorded — see the COI note on Typed Decision Verifiers. The ecosystem around it — ZTC, the open reproduction open-jev, Laya — exists because Jev named the category.

In a partner's orchestration (LangChain, 2026-09-25)#

The second first-party account is not TypeSafe's but an integration partner's: Building Prod with Jev and LangGraph (Sydney Runkle & Hunter Lovell, LangChain blog, vendor-claim — LangChain sells the LangGraph runtime and LangSmith tracing the post showcases). It adds three numbers, none from TypeSafe, all unmethodologized:

• Discovery review, 5–6× faster than Sonnet on the classification step. One LangGraph graph asks Jev three questions per litigation page in one request (responsive? contains PII? possibly privileged?), each answer mapping to a route — set aside, LLM redaction, or an attorney_review interrupt that pauses the graph for a human. Swapping Jev for Sonnet as the classifier on the same graph made the classification step 5–6× slower "across trials." No accuracy, agreement, page count, or Sonnet version is reported — speed only.
• Judge stability across 100 repeated runs. In "an early Jev-as-a-judge experiment," its scores "barely moved across 100 repeated runs, far less than any LLM judge we tested." No numbers, task, or judges named. This is run-to-run consistency, a different property from calibration or accuracy — a deterministic function is perfectly consistent and can be consistently wrong.
• Browserbase Stagehand act(): median latency 1.97 s → 0.46 s (~4.3×). Stagehand marks the page's interactive elements; Jev picks the action type and the best candidate element; any decision under 0.7 confidence falls back to an LLM. "Early testing," Browserbase's figure reported secondhand; no task-success or fallback-rate numbers.

The post's framing — Jev "takes one capability out of the frontier LLM bundle, judgment, and makes it a primitive too cheap to measure," deployed as "cheap by default, frontier on exception" (Jaya Gupta's "Great Unbundling of Intelligence") — makes Jev the first stage of a confidence-thresholded cascade rather than an LLM replacement: "Jev doesn't fully replace an LLM for most use cases." It restates TypeSafe's benchmark as "up to 200× faster and 400× cheaper" (the launch's 193.6× / 444.6×, rounded). One uncorroborated developer quote — "I'm basically converting every agent we have right now into a workflow powered by Jev" — is the only adoption signal.

Connections#

• Typed Decision Verifiers — defines the category; that page carries both the launch claims and the independent board that tests them
• Trained Calibration — RLCD is a vendor-named fourth route to trained calibration: calibration as the whole objective of a decision model, not a honesty add-on to a chat model
• The Bitter Lesson — the vendor's counter-slogan, "the bitterest lesson": optimizing for the right task matters more than data, compute, or algorithms
• Crystallizing Agent Work into Workflows — Jev's pitched use ("smart if-statements" inside hand-written control flow) is that page's Type 2 hybrid execution sold as a model product; the LangChain/LangGraph post is that Type 2 shape built directly, with no promotion path behind it
• Reasoning–Acting Interleaving (ReAct) — Stagehand's act() rebuild is that page's "action selection as classification over an enumerated valid set" shipped in production, with Jev's 255-way choice cardinality as the enumeration's ceiling
• Cost-per-Task Over Cost-per-Token — "cheap by default, frontier on exception": Jev as the cheap first stage of a confidence-thresholded cascade

Sources#

• Introducing System One Models & Jev — Diogo Almeida, TypeSafe AI blog, 2026-09-15, vendor-claim: comparison table, workflow-eval Pareto chart, structured-output / tool-call error-rate charts, FAQ. Chart values transcribed from images at ingest; scatter values approximate (log axis)
• Inside the JEV Ecosystem: 13 Answer Verifiers on One Test Set — Proto_AGI (mayafree), HuggingFace, 2026-09-20, empirical: the independent AUC board, retry-gate experiment, and hosted-API latency comparison
• Building Prod with Jev and LangGraph — Sydney Runkle & Hunter Lovell, LangChain blog, 2026-09-25, vendor-claim (integration partner; sells LangGraph/LangSmith): the discovery-review graph and its 5–6× classification speedup vs Sonnet, the 100-run judge-stability claim, Browserbase's Stagehand act() latency (secondhand), the three-ways-to-build figure and LangSmith decision-view screenshots (read at compile; contains_pii = 0.98 on the sample page)