The news
TypeSafe AI's Jev decision model, opened for early access on September 15, picks from answers a developer sets instead of writing text, and TypeSafe says it is up to 445 times cheaper than language models. By September 29, open-source clones could run locally.
Jev is the first public model in what TypeSafe calls System One Models, built for decisions inside software. A developer supplies context and questions, and Jev returns yes-or-no answers, choices from a fixed list or scores, each with a probability, according to Vercel's description. TypeSafe lists a price of $42 per billion input tokens, and output is not charged. Its website says Jev produces zero hallucinations; The Rundown noted that the guarantee covers the format of an answer, not whether the judgment is right.
TypeSafe says Jev is up to 194 times faster and 445 times cheaper than language models in its own workflow tests, according to Vercel. The Rundown reported that the company was founded by former OpenAI researcher Diogo Almeida. On September 18, Vercel said nearly 13% of paid teams on its AI Gateway were using Jev within 24 hours of launch, twice the share of OpenAI's GPT-5.6 family and more than six times that of Anthropic's Claude Fable 5.1. Vercel added that the next test is whether that early adoption lasts.
Open-source developers moved quickly. A GitHub project called Jeff, from developer firelex, offers fine-tuned open models based on Qwen3.5 and Gemma 4 that accept Jev's request format and return a decision in about 22 to 28 milliseconds, its README says; it states it is not affiliated with TypeSafe. PostHog (the product-analytics company) published Jeeves, a 9-billion-parameter decision model that adds a reasoning step, and says it scored 0.889 accuracy against Jev's 0.857 on held-out test data.
On September 29, The Register reported on Jevstiller, an open-source tool that trains a local model on Jev's past answers so familiar requests run on a device's own processor, in as little as 15 milliseconds against about 300 milliseconds for Jev. Anything the local model cannot handle still goes to Jev. Its developers target 98% agreement with Jev and warn that "agreement is not accuracy."
Not everyone is convinced the idea is new. On The Register's Kettle podcast on September 28, staff noted that classifiers, embedding models and cross-encoders already do similar work, that Jev can only choose among answers a developer supplies, and that TypeSafe's backend is proprietary.
The numbers
- Jev list price
- $42 per billion input tokens; output not charged
- TypeSafe's claimed cost advantage vs language models
- Up to 445x cheaper (company tests)
- Vercel AI Gateway paid teams using Jev within 24 hours
- Nearly 13%
- Jevstiller local response time vs Jev
- As little as 15 ms vs about 300 ms
- PostHog Jeeves vs Jev accuracy (PostHog test)
- 0.889 vs 0.857
Why CEOs should care
For CTOs and CFOs, the case is cost. Many enterprise AI calls are not creative writing but routine choices: route this email, approve this invoice, flag this alert. If a decision model handles those for a fraction of what an LLM costs, the savings could be large. Ask your teams what share of current LLM spending goes to classification and routing, then test a decision model against that slice. Treat TypeSafe's 445x figure as a vendor benchmark until your own workloads confirm it.
For CISOs and privacy leads, the local options may matter more than the price. Open models like Jeff and tools like Jevstiller can keep routine decisions on company hardware, which limits how much customer data leaves the building. Ask which sensitive workflows could run on a local model, and who will own accuracy testing, since a distilled copy inherits the original's mistakes.
For anyone negotiating a multi-year AI platform deal, this argues for flexibility. A contract priced on large LLM token volumes could look expensive if routine calls move to a cheaper decision layer. Ask for terms that do not penalize shifting workloads to other models, and avoid minimum-spend commitments sized to current usage.
The bigger picture
Jev fits a broader move away from one giant model doing everything toward a mix of specialized models matched to tasks. Its fast uptake on Vercel shows developers want cheap, predictable components for automation. The quick arrival of open-source clones shows how hard it is to protect a new model category once its request format is public. For enterprises, the likely result is more choice for simple AI tasks, and possibly more pressure on LLM prices for that work.
What’s next
Watch whether Jev's early adoption holds, as Vercel itself flagged; whether TypeSafe widens access beyond early access and publishes independent benchmarks; and whether open alternatives such as Jeff and Jeeves hold up on real enterprise workloads.
What “Fact-checked” means
Fact-checking means testing a story’s facts against the evidence before it is published. This story went through at least two separate checks before this version was published.
- What we checked
- Its names, figures, dates, job titles, quotes and who said what were checked against the story’s sources, including its main source where it could be opened. The headline was checked for accuracy and overstatement.
- How
- A first check reviewed the whole story. If it passed, a second, skeptical check went back to the sources to look for mistakes in the most important facts. If a check flagged the story, it was edited to fix the problems found, and a separate re-check then reviewed the whole story again.
- Who
- The checks are made by our newsroom, as steps kept separate from the writing, under rules set by our editor, Hussein Mukhtar. A story the checks still flag is held for the editor, who decides whether it is fixed, published or dropped.
- If something is wrong
- “Fact-checked” does not mean error-free. If a material error is found after publication, we correct the story and add a note saying what changed. Report an error
Companies in this story
