Martin Casado: Labs Missed Jev Due to the 'Presence' of Conversational AI
Martin Casado, a partner at a16z, stated that cutting-edge labs missed Jev because they focused on the 'presence' of conversational AI, while traditional software requires models that can make choices. Large models are based on chat, inputting text and outputting text, which is awkward when forced into programs; Jev selects the optimal option from given choices, making it faster, cheaper, and trainable for specific tasks.
Jev was launched by TypeSafe AI on September 15, with founders including former OpenAI researcher Diogo Almeida. The product does not generate prose but returns typed decisions: selecting from fixed options, scoring, or providing judgments and probabilities. It claims a delay of about 70 to 500 milliseconds, costs approximately $0.042 for inputting one million tokens, and does not charge for output; company evaluations indicate that specific workflows can be about 194 times faster and 445 times cheaper than large models.
Data from Vercel AI Gateway shows that around 13% of paid teams accessed it within 24 hours of launch, which is about twice the rate of the GPT-5.6 family and six times that of Fable 5.1, making it the fastest adoption in the history of the gateway. A waiting list cleared 140,000 people in about 36 hours, and it opened on September 20 with a $5 credit. The name is derived from the Jevons Paradox, emphasizing intelligence per dollar. Approximately $40 million in funding has been disclosed.
Casado framed the lab's goal as creating a deity: a god that speaks natural language. Jev targets if statements, sorting, and enumerating branches, with consumers being code rather than readers. In a podcast, he discussed with Aaron Levie and Steven Sinofsky how to separate decision engines from chatbots, stating that innovation will occur more in the software layer surrounding models.
The speed of adoption reflects a backlog: classification, routing, guardrails, and ticketing do not require paragraphs. Using generative models as classifiers incurs costs and delays on unnecessary text. Jev extracts this layer from prompts.
In market mechanisms, buyers are development teams that need to write millions of small decisions into existing programs, while sellers are model companies repackaging classifiers into billable APIs; funding flows from the gateway and token billing to dedicated decision reasoning, rather than flagship conversational models. Beneficiaries are the orchestration layer and gateway, while those under pressure are the bills from routing with large models. Event-driven factors stem from delays and unit prices, rather than new conversational capabilities.
No on-chain data. What can be verified is the gateway's share and pricing, not long-term retention.
Source: Public information
ABAB AI Insight
Casado uses a Nicira-style infrastructure intuition to dissect the model market: chat is the human-machine interface, while software requires branching. Labs have poured all their capital into conversational general intelligence, while classification, scoring, and routing are still temporarily commandeered by prompts. Jev has productized that commandeered layer, which is why 13% of teams appeared on the gateway in one day — that’s not a new belief, but an old bill finding an outlet.
The capital path is to narrow tasks cut from former OpenAI training methods, using low-cost input tokens for high-frequency calls. The $40 million level of funding is not aimed at creating a deity, but at turning if statements into cloud services. Vercel and OpenRouter are competing for access points: whoever first integrates the decision model into default routing will siphon off classification traffic. The Jevons naming suggests that cheaper decisions will lead to more decisions, and the total billing may not necessarily decrease.
Similar structures appear in search, separating sorting from page generation, in payments, separating risk control from customer service dialogues, and in CDNs, separating caching from origin computation. The industry position is at a division after the inflation of general models: those who can write continue to write, and those who can choose specialize in selection. If labs insist on everything being conversational, the software layer will end up nurturing classifiers on its own.
Structural judgment belongs to technological substitution. The mechanism is that dedicated decision models replace misused generative models to execute program branches. Once substitution is established, the unit of pricing intelligence shifts from 'a paragraph' to 'a choice'; the race to create deities and software engineering will separate again.
ABAB News · Cognitive Laws
Speaking models are not necessarily the models that software needs.
Once options are given, generating text becomes wasteful.
While labs are creating deities, programs are still looking for components that can make choices.