Typesafe opens Jev to the public with $5 credits as builders wire it into sub-second bots
The probability-only model returns typed odds in 70–500ms, and early BTC strike outputs diverge from Polymarket’s $67.8M market.
Typesafe AI has opened its System One model “Jev” to the public with no waitlist and $5 in credits per new user, making it cheap to test in automated trading and scoring pipelines. Early crypto experiments are already pushing Jev’s 70–500ms probability outputs into sub-second onchain execution loops, while a BTC strike test produced odds well below Polymarket’s implied probabilities.
Jev Goes Public: Probability-Only AI Built for Trading Loops
Typesafe AI removed the waitlist for its System One model, Jev, and paired the public opening with $5 in credits for each new user. The team’s announcement was blunt: “Jev is now available to everyone. No waitlist.” That combination matters less as a marketing beat than as a workflow change, because it turns Jev from a scarce demo into something builders can hammer on in production-like loops without asking permission.
Jev’s core mechanic is narrow by design. It does not generate prose or run a chat interface. Users provide a state plus a closed question, meaning the answer space is bounded to fixed options like yes/no, buy/sell, or a numeric score, and Jev returns typed probabilities in about 70–500 milliseconds.
The pricing model is also built for repeated calls. Typesafe priced Jev at $0.042 per million input tokens, with output free. Input tokens are the units of text or data sent into the model that determine usage cost, so the economics tilt toward frequent, small queries rather than long responses.
Typesafe has worked on Jev for two years in stealth and raised $40 million. The pitch is infrastructure-style AI: a component that can be called by other software, where the output is a confidence-weighted decision rather than a paragraph that needs interpretation.
Early Crypto Integrations: A ~300ms Onchain Bot Posting on Kuru
The cleanest crypto-native example so far is an execution loop, not a dashboard. On Sept. 16, Jarrod Watts, lead AI engineer at Monad, built and open-sourced a Jev-driven onchain trading bot that consumes a price feed, asks Jev “buy or sell,” and then posts orders on Kuru.
Watts described the mechanism directly: “Jev decides if it should ‘buy’ or ‘sell,’ given the price feed of an asset pair, and executes real trades.” The bot “rest[s] a post-only limit on Kuru’s onchain book about every 300 milliseconds,” and the code is published as jarrodwatts/jev-trader.
A post-only limit order is a maker-only instruction that must add liquidity to the order book and will be canceled if it would immediately execute as a taker. On an onchain order book, that maker-only constraint is doing real risk work, because it forces the strategy to live or die on getting filled rather than crossing the spread every loop.
The broader point is that Jev’s output format fits directly into execution code. A chat model can be fast, but it still returns text that needs parsing and guardrails. Jev returns a typed probability on a bounded action, which is why builders are wiring it into sub-second loops instead of treating it like an assistant.
Jev vs. Polymarket on BTC 2026 Strikes—and What That Divergence Signals
A second early test used Polymarket as a benchmark for crowd-implied probabilities. On Sunday evening, Jev was fed Polymarket’s live event data for “What price will bitcoin hit in 2026?” and asked a series of yes/no questions framed around whether a Binance 1-minute candle would print specific strikes before 2027. A Binance 1-minute candle is the one-minute OHLC bar from Binance, used here as the reference for whether a level was reached.
Polymarket’s pricing is expressed in cents on the dollar, where 81 cents implies roughly an 81% chance. In that market, traders were pricing the $85,000 strike at about 81%, with event volume near $67.8 million. The same test had Jev assign 55% to $85,000, 48% to $90,000, and 34% to $100,000.
The context matters because this was not a thin market. The event had already seen $80,000 “printed,” and $75,000 and $65,000 “tagged,” which makes the $85,000 line a near-term continuation bet rather than a pure long shot.
The divergence does not prove mispricing. The test description does not include prompt formatting, model settings, or a reproducible harness, and there is no live PnL attached to the Jev outputs. What it does suggest is that Jev can generate materially different priors than a liquid crowd market, which is useful as a cross-check if the methodology can be made repeatable.
My Take: Jev Is a New Input, Not a New Edge—Yet
The threshold that matters here is not “AI in trading,” it’s whether probability-only outputs at 70–500ms actually compress the signal-to-execution path enough to survive fees, slippage, and adverse selection. Jev being open-access with $5 credits and cheap input pricing makes it frictionless to prototype, and the open-sourced ~300ms Kuru loop lowers the barrier for copycats to iterate fast.
The real test is whether Jev-based repos start publishing verifiable metrics like PnL, fill rates, and slippage, and whether the Jev vs. Polymarket gap persists across multiple high-volume events with reproducible prompts. If those numbers show up and hold across regimes, Jev becomes a practical building block for onchain execution rather than just a novel way to ask the market a question.
