TypeSafe Jev draws lookalike resellers and an OpenAI preview

TypeSafe’s Jev decision model, priced at $0.042 per million tokens, drew fake resellers and a similar OpenAI Decisions API, plus new accuracy and adoption figures.

TypeSafe AI released Jev 1.13 on September 15, 2026, a closed decision model that returns typed Choice, Score, and Noul—or, in a later explainer, Bernoulli yes-or-no—probabilities calibrated by reinforcement learning instead of generated text. Diogo Almeida, described as founder by one outlet and CEO by another and a former OpenAI engineer, sells it through TypeSafe, Vercel AI Gateway, and OpenRouter at $0.042 per million input tokens; vendor figures cite up to 193.6x speed and 444.6x lower cost versus GPT-6 Astra and Fable 5.1, roughly 70–500 ms latency, a 32K window, free output, and later 67.8% mean workflow accuracy at about $0.0004 per case. Eye Security’s Dion Fieret and Lucas Hop reported fake storefronts, including jev-ai.pro, that forward prompts to the genuine API while charging about $0.247–$0.483 per million tokens, or roughly 6–11.5 times list price. GBHackers said about 670 domains containing “jev” received TLS certificates within eight days of launch, while Cyber Security News said lookalikes were registered within three days and that jev-ai.pro ranked first on Google; both describe a six-site cluster involving Namecheap and Cloudflare, with one Railway app also mentioned. The outlets agree on overcharging and third-party data-exposure risk, but Cyber Security News said no malware or confirmed prompt theft was established, whereas GBHackers also flagged possible API-key theft. TechCrunch reported that OpenAI CEO Sam Altman previewed a limited Decisions API constraining Luna to predefined choices, which Almeida likened to a clone, alongside a QueryStory demo of about $2.94 versus $372 per action check; an October 3 explainer added Fastino’s GLiDE, GLiNER2.5-Decide, open-source rivals, and Vercel’s report that nearly 13% of paid AI Gateway teams used Jev within 24 hours.

• TypeSafe AI released closed Jev 1.13 on September 15, 2026, via TypeSafe, Vercel AI Gateway, and OpenRouter; weights are unreleased, input is $0.042 per million tokens, output is free, the window is 32K, and latency is about 70–500 ms.
• Jev returns Choice (up to 255 options), Score, and Noul or Bernoulli yes-or-no probabilities calibrated by reinforcement learning (RLCD), not generated text; questions can run in parallel on one state.
• Vendor evals claim up to 193.6x speed and 444.6x lower cost versus GPT-6 Astra and Fable 5.1; a later workflow eval reported 67.8% mean accuracy, about $0.0004 per case and 0.4 seconds, matching Claude Sonnet 5 and trailing a top model at…
• Eye Security’s Dion Fieret and Lucas Hop found fake shops, including jev-ai.pro, that proxy prompts to the real API at about $0.247–$0.483 per million tokens (roughly 6–11.5x); a six-site Namecheap/Cloudflare cluster was described, one hop…
• Timing and harm claims differ: GBHackers said about 670 “jev” domains received TLS certificates within eight days and flagged possible API-key theft; Cyber Security News said lookalikes were registered within three days and that no malware…
• OpenAI CEO Sam Altman previewed a limited Decisions API that constrains the Luna model to set choices; Diogo Almeida—called founder by one outlet and CEO by another, and a former OpenAI engineer—likened it to a clone. A QueryStory demo…
• Vercel said nearly 13% of paid AI Gateway teams used Jev within 24 hours. Named rivals include Fastino GLiDE, GLiNER2.5-Decide, Laya, kev, and open-source reproductions.