37x Valuation in 24 Days: How Decision Model Jev Ignited Silicon Valley's Newest AI Category

In just 24 days, an AI company that had been virtually invisible in Silicon Valley pulled off a textbook-worthy reversal. According to Quantum Bit, Jev's parent company TypeSafe AI announced an $870 million Series A round, pushing its post-money valuation to $7.5 billion — a 37.5x surge from the $200 million seed-round valuation in mid-September, with the funding amount itself growing nearly 22x.

What makes this even more striking is that the product behind all the hype isn't yet another large model that can chat or write code. It's a decision model that "only judges, never speaks." It strips away the generative capabilities that traditional LLMs are most proud of, and focuses on doing exactly one thing: making decisions.

Why Is a Model That Only Says Yes or No Worth $7.5 Billion?

Jev's product positioning is deeply counterintuitive. It doesn't chat, doesn't write code. When faced with a yes-or-no question, it chooses "or" — meaning it handles classification, scoring, and selection, without producing long-form text output.

Precisely because the architecture is lean and the cost is almost immeasurably low, TypeSafe AI made output tokens permanently free. Input pricing is just $0.042 per million tokens — low enough to put serious pressure on the Flash-series lightweight models from major providers.

Capital doesn't chase social media buzz alone. What truly convinced investors was the growth curve. According to Vercel's official blog, within 24 hours of Jev being integrated into its AI Gateway, nearly 13% of paying teams had adopted it — more than twice the adoption rate of any previous model launch, making it the fastest-adopted model in AI Gateway history. a16z disclosed that within just 3 days of launch, Jev had processed over 1 trillion tokens.

By the time the Series A was announced, TypeSafe AI reported that roughly one-third of Fortune 500 companies had begun using or integrating Jev. a16z, Sequoia Capital, DCVC, and other top-tier firms all joined the round, with a16z general partner Martin Casado — widely regarded as the godfather of infrastructure investing — personally joining the board.

What is a Decision Model? The core difference between a decision model and a traditional large language model lies in the design of the output space. LLMs produce open-ended natural language sequences, generating tokens one by one, with computation costs scaling linearly with output length. Decision models constrain their outputs to a finite set of structured results — such as category labels, scores, or boolean values — resulting in shorter inference paths, lower latency, and outputs that downstream software can consume directly without additional parsing. This design is particularly well-suited for machine-to-machine (M2M) call patterns: tasks like request routing, content moderation, priority ranking, and risk scoring are fundamentally about "making a choice" rather than "generating text," and decision model architecture maps naturally onto these needs. By pricing output tokens at zero, Jev directly translates its architectural advantage of "no generation needed" into a pricing strategy.

Backend Calls Are the Real Moat

The logic behind Jev's explosive growth is fundamentally different from how chatbots go viral.

Chatbot traffic typically comes from curious humans trying something new — the novelty fades after a few uses. Jev's calls, by contrast, happen overwhelmingly in the background: once software integrates it, it runs around the clock classifying, scoring, and making decisions.

CEO Diogo Almeida said in a Latent Space interview that the company's daily processing volume had surpassed one trillion tokens — and not as a one-time peak. Overnight traffic continued to climb. His line, "People sleep, but code doesn't," captures the commercial thesis perfectly: demand comes from machine automation, not human curiosity, which means it's continuous and scalable.

This may be exactly what investors are most excited about. A "component" embedded in production systems, called millions of times a day by software, offers far more predictability than a chat app living and dying by traffic spikes. In its funding announcement, TypeSafe AI kept things characteristically straightforward — no lofty promises, just a statement that it plans to build more machine-native models, flesh out the infrastructure intelligent software needs, and ship the features enterprise customers have long been asking for.

What is AI Gateway? AI Gateway is a unified access layer offered by cloud platforms like Vercel, allowing developers to route, monitor, and bill API calls across different AI models — such as OpenAI, Anthropic, and various open-source models — without needing to integrate each provider's SDK separately. For assessing Jev's true adoption speed, AI Gateway data carries high credibility: it measures paying teams actively integrating Jev into production workflows, not free trials or web demos. This makes "13% of paying teams adopted within 24 hours" a meaningful indicator of developers' judgment about real-world value, rather than mere curiosity.

The Man Who Taught AI to Talk Is Now Telling It to Shut Up

The most dramatic irony surrounding Jev comes from its creator, Diogo Almeida.

Almeida previously worked at Google Brain and OpenAI, serving as a core author on the InstructGPT paper and contributing to work on RLHF, ChatGPT, and GPT-4. In short, he was one of the people who taught large language models to "talk nicely to humans."

But after leaving OpenAI to start his own company, he became increasingly convinced that the entire industry was pouring too much energy into "making AI speak." Over the two years building TypeSafe AI, he kept returning to one question: models are already smart enough to write code, perform complex reasoning, and even crack hard math problems — so why do most critical steps in real-world workflows still require human confirmation?

His answer: today's AI is first trained to be a "people-pleasing assistant," not a component that software can reliably call. It's optimized to make responses feel good to humans, not to run in the background of a system millions of times a day.

This explains TypeSafe AI's distinctively memorable tagline: "Build Prod, Not God." Almeida has been openly critical of the industry's obsession with benchmark competition, arguing that public benchmarks are trivially easy to game — and that even without deliberate manipulation, developers can end up chasing leaderboard rankings instead of real utility. So when Jev launched, there were no crushing benchmark tables. Instead, he showed what it actually does and let developers drop it into their workflows to see if it could get the job done.

The team behind it is equally formidable: COO Sasha Sheng comes from Meta/FAIR, CTO Erik Gafni is a serial entrepreneur, and the broader team brings together veterans from OpenAI, Google Brain, and Meta/FAIR. Together, they're betting on a path diverging from "keep scaling bigger models" — the idea that AI may already be smart enough, and what's really missing is an interface that software can trust.

What is InstructGPT? InstructGPT was a landmark model released by OpenAI in 2022, whose core contribution was applying RLHF (Reinforcement Learning from Human Feedback) to language model alignment at scale. The approach works by having human annotators rank multiple model responses, using those preference signals to train a reward model, and then fine-tuning the language model with that reward signal so its outputs better match human expectations. InstructGPT is the direct technical predecessor of ChatGPT and the foundational work behind the entire conversational AI assistant wave. As a core author of that paper, Diogo Almeida was deeply involved in teaching AI to "answer questions in a way people like" — which makes his subsequent pivot to stripping out generation entirely and focusing purely on judgment all the more compelling. He understands the costs and limitations of conversational capability better than almost anyone.

In 24 Days, a New Category Faces Challengers From All Sides

If Jev's single-product breakout is only half the story, the other half is arguably more significant: it made the entire industry suddenly realize that "decision models" represent a viable new category. And once a category is validated, TypeSafe AI went from having almost no direct competitors to facing challengers from every direction — almost overnight.

The incumbents moved with striking speed. Just about two hours after Diogo announced the funding, Microsoft CEO Satya Nadella unveiled Microsoft-Decision-1, a new decision model built on post-trained Qwen3.5-9B, specifically designed to handle routing, classification, priority judgment, validation, and workflow control.

OpenAI moved even earlier. On October 6 — just 21 days after Jev's launch — it opened public beta access to the Decisions API, powered by GPT-6 Luna, capable of converting text and images into three types of structured outputs: Predicate, Choice, and Score. Even before that, on October 1, Cloudflare had already released open-source decision models Clef and Clef-flash, both compatible with the Jev API and explicitly positioning themselves against Jev in head-to-head comparisons.

All of this competition materialized within a span of 24 days. And that's precisely what makes this funding round truly valuable: Jev didn't just build a breakout product — it validated a thesis. AI doesn't have to get better at talking. It can get better at staying quiet and doing its job reliably inside software systems.

From an invisible startup to a $7.5 billion company drawing Microsoft, OpenAI, and Cloudflare into the same arena, TypeSafe AI used 24 days to prove there was a real gap in the market no one had filled. The question now: when the giants arrive in force, can the pioneer hold the territory it opened up?