AI has spent years learning to talk. Now a new class of models is learning when to choose, route, score, and act. Jev has turned those decisions into a major competitive battleground, and OpenAI has opened its own answer to customers through the Decisions API.
The shift matters because many software tasks do not need a paragraph, a brainstorm, or a long chain of reasoning. They need one reliable choice, delivered with a confidence score and little delay. That simple idea has pushed TypeSafe to a valuation of $7.5 billion weeks after launch, with another figure placing the company at $7.6 billion.
Jev Turns AI Into a High-Speed Decision Engine
Led by Diogo Almeida, a former OpenAI engineer instrumental in building ChatGPT, TypeSafe created a category for deep learning models that act as super classifiers. Jev uses the same transformer architecture behind ChatGPTs and the Claudes of the world, but it does not generate language.
Instead, Jev evaluates a query against predefined options, assigns each option a probability, and returns specific choices backed by confidence scores. Sathya AG describes decision models as “specialised AI models that pick a single, pre-defined answer based on a confidence score rather than generating a free-form response.”
That design targets work such as sorting support tickets, scoring candidates, routing requests, and choosing an AI agent’s next action. Sathya AG compares using an LLM for these deterministic workflows to “using a smartphone exclusively for its flashlight capability.” The argument is direct: when a task needs a fast, reliable signal, a model built for decisions can avoid the cost of language generation.
TypeSafe claims Jev can run up to 193.6 times faster and cost 444.6 times less than conventional LLMs for structured decision-making workflows. Hiring marketplace Jack & Jill replaced Gemini 3.1 Flash Lite with Jev for candidate scoring, reporting 88% lower costs and roughly half the median processing time, with no statistically significant change in measured quality.
Developers have already used Jev for Claude Code skill selection, as well as agents playing soccer and Super Mario. The most entertaining test may be JevMan, where decision models play PacMan. Jev tops that ranking, and Felix Wunderlich of Opper AI calls the benchmark “an interesting test because of its nonstop stream of quick, high-stakes decisions.”
OpenAI, Liquid AI, And Big Tech Chase The Same Opening
Jev’s momentum created an obvious question for the rest of the industry: why leave this market to one startup? Liquid AI introduced d1 as the first model it said could rival Jev on text-based decisions while extending the approach to images. Across six applications, Liquid AI said d1 matched or beat GPT-6.1 Sol on four, at 19 to 200 times lower cost and with lower latency across all six.
Two days later, Liquid AI released open-weight versions, including a 3-billion-parameter model designed to run from data-centre GPUs to edge devices. Cloudflare released Clef and Clef-flash as open-source decision models, AWS released Strands Decider 2B through Strands Labs, and Microsoft joined with Microsoft-Decision-1, post-trained from Qwen3.5-9B.
OpenAI’s response arrived at its developer day on September 29. The company announced a Decisions API that brings ChatGPT Luna’s capabilities to structured tasks such as classifying inputs, routing requests, and selecting an AI agent’s next action. OpenAI later made the model widely available to users for the first time, adding voice and visual modalities to the public version.
Nikunj Handa, product lead for OpenAI’s APIs, said colleagues saw Jev and thought, “wow, we we have like a million use cases for this.” His team reconfigured ChatGPT Luna to speed inference and lower costs, creating an API that shrinks response times from one to three seconds to 100, 120, or 180 milliseconds.
That speed unlock reaches into real products. In initial trials with Box, Datadog, Ramp, and Speak, engineers used the API to replace OpenAI’s LLMs with higher throughput. Andrew Hsu, CTO at Speak, said the time between actions fell from 1.4 seconds to .13 seconds, while OpenAI used the product to improve its Dots voice assistant’s responsiveness.
OpenAI is also targeting lower latency by placing inference resources geographically around the world. Handa sees an even faster future, asking: “Can we get this to something like you know 60 milliseconds, and then you could imagine like an API like call like for every single frame in like a video?”
Benchmarks, Safety, And The Bigger AI Bet
TypeSafe CEO Diogo Almeida distrusts benchmarks and urges teams to trial models in their own work. TypeSafe’s latest Jev model adds image support and smart-if statements, while public rankings send mixed signals: the AIM System One Index and Artificial Analysis Decision Index favor D1 Ultra, yet Jev remains competitive in seventh and third place, compared with OpenAI’s Decision Model in eighth and fourteenth.
TypeSafe’s offerings are currently the most widely queried on OpenRouter. Handa says Luna is “extremely powerful and has been showing better calibrations,” especially for image classification, while OpenAI customers want more flexibility around using images as outputs.
The decision-model surge also raises a harder question about how AI companies spend their resources. Jain says, “All the top AI companies do care a lot about safety, and they invest in it, but they could be investing a lot more,” adding, “If they invested more in safety, I would still be at the lab and working on what I’m interested in with a large team.” DeepMind’s principles state that it maintains “a holistic approach to responsibility, safety and security.”
Funding for AI safety is growing. METR says it raised $71 million in six months this year, while AI safety groups and individual donors spent nearly $50 million in this year’s election on candidates focused on regulating AI. Jaan Tallinn donated $170 million to nonprofits focused on AI safety, Coefficient Giving announced Project Tailwind in September, and the group says it is on track to give more than $1 billion this year. Sampura Research received an $11 million grant.
At the same time, developers and investors are still waiting for mature agentic products to turn this infrastructure into revenue. Ars OpenForum thread starter AndrewZ wrote, “What we haven’t seen yet are products/services based on mature agentic. I think this is the potential AI revolution.” AndrewZ also wrote, “I still see the stock market bubble popping, but when the dust settles, I think this is where AI companies will be earning revenue.”
That future will demand more than picking the fastest model today. Dan Haiem, founder and CEO of AppMakers USA, wrote that founders often treat model selection as a one-time technical decision, even though they are making an architectural commitment that can outlast a model’s changes, price shifts, or replacement.
Jev did not remove the need for AI fundamentals. It exposed a new layer of them: knowing when language generation is unnecessary, when confidence matters, and when a specialized decision model can make software faster, cheaper, and ready for the next wave of agents.
Based on
- Jev Isn’t An Overnight Fix For Your Lack of AI Fundamentals — analyticsindiamag.com
- OpenAI makes its Jev competitor widely available as decision model competition heats up | TechCrunch — techcrunch.com
- The AI safety bottleneck: Save the world or get the bag — africa.businessinsider.com
- The Generative AI Bubble Is really Going to Pop – Part Deux | Page 35 | Ars OpenForum — arstechnica.com
- The AI Model Built On Today Might Not Exist Tomorrow — forbes.com
