<- all tokdocs
2026.09.17@ducktate.com

Jev, TypeSafe AI's "System One" Model, Costs $0.042 Per Million Input Tokens and Cannot Write a Line of Code

AI
Watch on TikTok

View on TikTok ->

Jev is a real model with real early-access availability, and the reason it cannot code is architectural rather than a missing feature. In episode six of his Token Hacking series, @ducktate.com describes getting access to Jev, wiring it into Claude, and discovering it would not build features for Ducktate, his Mac note-taking app. TypeSafe AI positions Jev as the first public "System One" model, a class that returns typed probabilistic decisions instead of generated text. The frustration in the video is the correct first reaction, and the follow-up question about what it is actually for has a concrete answer.

What TypeSafe Actually Shipped

TypeSafe AI announced Jev in mid-September 2026. Sources disagree on the exact day: the news ticker shown in the video reads September 15, The Register's coverage is dated September 16, and the TypeSafe announcement post currently carries a September 18 publish timestamp.

The product documentation is specific about the shape of the thing. You send a block of unstructured state plus a set of typed questions to POST /v1/systemone, using the model alias jev-latest, and Jev evaluates all of the questions in a single parallel pass. What comes back is typed answers with confidence probabilities attached, for example a route label of "billing" with a confidence of 0.95, or a frustration score on a 0 to 2 scale. The docs state plainly that System One models "do not write replies, produce code, or generate explanations of their reasoning." Input is text only for now, with no image, audio, or video support.

That matches the list in the video almost exactly. Classify, route, score, rank, give confidence, pick a branch, select a tool call, and judge outputs are the documented use cases. The video's claim that it has no step-by-step reasoning stage is also consistent with the documentation.

The Pricing Is As Strange As He Says

The video's numbers hold up. TypeSafe lists input at $0.042 per million tokens, which it also phrases as $42 per billion input tokens, and output tokens at $0 with the note "too cheap to meter." Free output stops looking like a stunt once you understand the model's shape: the output is a small typed struct with a handful of probabilities, not a stream of prose, so there is almost nothing to meter.

TypeSafe's homepage headline benchmark shows a single workflow costing $0.000081 and finishing in 0.114 seconds on Jev, against $0.013880 and 8.566 seconds for an LLM baseline, which it summarizes as 193.6x faster and 444.6x cheaper. The video shows a comparison table with that same 193.6x figure for Jev at $0.39, alongside Claude Haiku 4.5 at 6.2x and $19.49, Claude Opus 5 at 2.1x and $176.05, Claude Sonnet 5 at 1.0x and $117.38, and GPT-5.6-Luna at 6.0x and $3.31. TypeSafe's written announcement cites comparisons against GPT-6 Astra and Fable 5.1 instead, so the on-screen table appears to come from a configurable widget rather than a fixed published benchmark. Elsewhere TypeSafe gives a broader range of 70 to 500 milliseconds end-to-end and 40x to 200x faster than frontier LLMs.

Every Number Comes From TypeSafe

This is the part the video leaves out, and it matters more than any of the speed multipliers. At the time of the announcement there was no third-party benchmark suite, no published paper, and no open access that would let outside researchers run their own tests. TypeSafe gates access behind a waitlist and labels the product Version 0.01 in its own footer. The latency figures, the cost figures, and the claimed 0% structured output error rate are all first-party.

The "cannot hallucinate" claim needs the same caution. It is true in a narrow, checkable sense: a model that emits a value from a fixed enum cannot invent a category that is not in the enum, and it cannot produce a type error. It says nothing about whether the chosen category is correct or whether the attached confidence number tracks real-world accuracy. RLCD, or Reinforcement Learning for Calibrated Decisions, is TypeSafe's proposed training alternative to RLHF, and the pitch is that it optimizes for a decision plus an honest probability rather than for an answer a human rater approves of. Whether that calibration holds up under independent testing is still an open question.

How He Got It Into Claude

The video skips the mechanics, and they are worth knowing. TypeSafe ships client SDKs and an HTTP API. Connecting Jev to Claude Code or Claude Desktop goes through MCP, and the connectors doing that work are community projects rather than official TypeSafe releases. Searching GitHub turns up a scatter of independent repos with names like typesafe-mcp and jev-mcp, several of which package themselves as Claude Code plugins and expose Jev as typed tools for classify, score, check, and batched ask. Reports describe more than 300 Jev-related repositories appearing on GitHub within two days of the announcement.

The practical consequence is that Claude keeps doing the writing and Jev handles the judgment calls inside the loop. Claude drafts the code, and Jev answers questions like "is this diff risky" or "which branch should this request take" for a fraction of a cent, with a probability the surrounding code can threshold on.

Where the Video Oversells

The closing line, that Jev will "reimagine every single workflow that uses AI right now," goes further than anything TypeSafe or its coverage supports. A more grounded version of the same point: workflows that currently burn a frontier model on a yes-or-no question are paying for text generation they throw away, and those specific calls have an obvious cheaper substitute if Jev's accuracy claims survive scrutiny.

The video also skips TypeSafe's most interesting demo. The Register reported on a Doom-playing bot driven by Jev reading structured game state rather than screen pixels, running roughly 10 queries per second at about $7 per hour. That demo argues the case better than any benchmark table, because a game loop is exactly the high-frequency repeated-decision workload the model is built for. The Register also reported $40 million in funding for TypeSafe and that CEO Diogo Almeida previously worked as a researcher at OpenAI.

Key Takeaways

  • Jev is a "System One" model from TypeSafe AI that returns typed decisions with calibrated confidence scores, with no text generation, no code generation, and no chat interface.
  • Pricing of $0.042 per million input tokens with free output is confirmed on TypeSafe's own site, matching the video's "four cents" figure.
  • The right mental model is a cheap judgment layer inside an existing agent loop, handling classification, routing, scoring, branch selection, and output judging while a conventional LLM does the writing.
  • Nothing about the performance claims has been independently verified. The 193.6x speed figure, the 444.6x cost figure, the 70 to 500 millisecond latency range, and the 0% structured-output error rate are all TypeSafe's own numbers, with the model behind a waitlist and no published paper.
  • The comparison table shown in the video, including the Claude Haiku 4.5, Opus 5, Sonnet 5, and GPT-5.6-Luna rows, does not match the model comparisons in TypeSafe's written announcement, so treat those specific figures as unconfirmed.
  • "Cannot hallucinate" holds only for output format. A model picking from a fixed enum cannot invent a label, and it can still pick the wrong one.
  • The Claude integration runs through community-built MCP servers, not an official TypeSafe plugin.

Resources

Published September 17, 2026. Writeup generated from a favorited TikTok.