What Is Jev? TypeSafe's System One Model and Its Pricing Claims
By Matt Cullerton, founder of Mavric and Apso
Jev is a model from TypeSafe AI built to make fast, structured decisions for software, such as picking an option or scoring against a rubric, instead of generating text. A TikTok from @promptly_media shows TypeSafe CEO Diogo Almeida explaining it, and everything he says matches TypeSafe's documentation. The pricing claims on TypeSafe's homepage are accurate in one case, compare against the most expensive model available, and in another case do not match the numbers printed beside them.
What is Jev?
Almeida says in the clip: "Jev is our first large programmable model or System 1 model, whatever you want to call it." TypeSafe's homepage calls Jev its "first public System One Model, optimized for automation," and the docs describe it as designed for "fast, structured decisions that software can use directly."
The API has three primitives:
- Choice returns a selected option with probabilities and a confidence value.
- Score returns a rubric-based score the same way.
- Noul returns a 0 to 1 truth value for a statement.
There is no text generation path. Almeida says Jev is "optimized for intelligence per dollar" and names the trade-offs as "reliability, cost, calibration, and speed." On the Latent Space podcast he argues calibration matters more than determinism, and every Jev primitive returns a confidence value with its answer.
The name refers to the Jevons paradox, after the economist William Stanley Jevons. Wikipedia's entry says Almeida chose it to suggest that cheaper machine intelligence will be used far more widely. The clip's auto-captions misspell it as "Jevin's Paradox" and the model as "Jeff."
Who created Jev?
The TikTok calls Almeida "The Creator of Jev." He is TypeSafe's co-founder and CEO, and SiliconANGLE names Erik Gafni and Sasha Sheng as the other two founders. SiliconANGLE reports Almeida previously worked at OpenAI on reinforcement learning from human feedback, InstructGPT, ChatGPT and GPT-4. Jev is trained with what TypeSafe calls Reinforcement Learning for Calibrated Decisions, an alternative to RLHF.
Jev is not open source. Wikipedia lists it as proprietary, with architecture and weights unpublished, and TypeSafe offers docs and a hosted API but no source code.
TypeSafe exited stealth on 2026-09-15 with a $40 million seed led by DCVC, at a $200 million valuation per Forbes as reported by SiliconANGLE, and released Jev in limited early access the same day. General availability followed on 2026-09-21. The current stable version is jev-1.13.0.
The TikTok itself is a horizontally mirrored cut of the Latent Space episode with Almeida, published on 2026-09-21, the day before the TikTok was posted. It shows only Almeida talking and no demo of the model.
How much does Jev cost?
TypeSafe's homepage lists "$42 Per Billion input tokens," which is $0.042 per million, and "238x Lower input price than Claude Fable 5.1." Anthropic lists Claude Fable 5.1 input at $10 per million. Ten divided by 0.042 is 238.1, so the math is exact.
That comparison uses the top of Anthropic's price list. Anthropic also lists Claude Haiku 5.5 at $0.10 per million input tokens for prompts up to 100,000 tokens, and Wikipedia reports Jev's input limit as 64,000 tokens, inside that tier. Against Haiku 5.5, Jev's input price advantage is 2.4x.
Jev's output tokens are free, but its outputs are only a choice, a score or a probability, so that follows from the design. TypeSafe also says it cannot yet show its pricing is unsubsidized and expects costs to fall.
Is Jev really 193.6x faster?
The homepage shows "193.6x Faster, 444.6x Cheaper" above a comparison of 0.114 s against 8.566 s and $0.000081 against $0.013880. Those figures give 75.1x faster and 171.4x cheaper, not the headline numbers.
TypeSafe's launch post explains the gap: "This is where the claims of 193.6x faster, 444.6x cheaper on our home page come from." The headline averages four workflows measured against "the average of the smartest models (in this case, Astra and Fable)," and the homepage shows only the simplest. The post also says the workflows "were made by individuals on our model capabilities team, some bias could exist." The baseline is TypeSafe's own System One wrapper around an LLM.
Independent results are more mixed. A survey of third-party audits reports that Every measured 777 judgments in under 0.7 seconds for roughly a quarter of a cent, about 25x faster than Claude Fable 5.1, while catching 6 of 7 planted defects against Fable's 7 of 7. The same survey cites a calibration study that found an out-of-distribution expected calibration error of 0.107 against a 0.024 noise floor, while Jev stayed between 0.024 and 0.032 on public benchmarks. If that holds, Jev's confidence values are reliable on some tasks and less so on others.
What are Jev's limits?
TypeSafe publishes a "jaggedness" page for jev-1.13, last reviewed 2026-10-02. It says the model "does not count reliably," "reads dates as text, not as ordered quantities," and "answers the question you wrote, not the one you meant." It adds that instructions with double negatives are answered less reliably and that "accuracy falls as the state grows with content unrelated to the decision."
Those limits matter for SEO and audit claims circulating about Jev. Ryze AI reported one such claim from Distribb's founder and states that it "has not been independently reproduced, and Ryze AI has not run it."
Bottom line
Jev is a real, proprietary model for structured decisions, and Almeida's description matches TypeSafe's docs. The 238x price claim is exact but compares against Anthropic's most expensive model, and Haiku 5.5 brings the gap to 2.4x. The 193.6x speed headline comes from TypeSafe's own four-workflow test, so check it against your workload, especially for counting or date tasks.
Sources
- Latent Space episode with Diogo Almeida: source of the footage.
- TypeSafe AI homepage: pricing and speed claims.
- Introducing System One models and Jev: how the headline multipliers were measured.
- TypeSafe AI documentation: Choice, Score and Noul primitives.
- Jev 1.13 jaggedness: documented limits.
- Anthropic pricing: Fable 5.1 and Haiku 5.5 input prices.
- Jev on Wikipedia: release, licensing, input limit and name origin.
- SiliconANGLE on TypeSafe's stealth exit: funding and founders.
- Jev for SEO (Ryze AI): an unreproduced SEO audit claim and its disclaimer.
- Jev audits roundup (OrcaRouter): general availability date and calibration findings.
Published September 22, 2026. Based on a TikTok by @promptly_media. Full video notes (markdown)