Overview
Yupp let people compare responses from AI models and express a preference. It offered access to models at no charge and rewarded feedback, then sought to sell aggregated evaluation data to AI developers. Founded by Pankaj Gupta and Gilad Mishne, the company launched publicly in 2025.[1] [2]
In March 2026 Yupp announced it was winding down. The founders said it had signed up more than 1.3 million users and gained several paying AI labs, but had not found strong enough product-market fit. Gupta also argued that AI use was shifting toward systems that combine models with tools and other services.[3] [2]
Image 1 / 1
Founding Story
Gupta and Mishne had worked together on consumer machine-learning products at Twitter, according to investor a16z. Gupta later led consumer engineering work at Google Pay and Coinbase; Mishne worked on machine learning at Google X. They founded Yupp in 2024 to connect people comparing AI responses with organizations evaluating models.[1] [4]
The company's premise had two sides. A person gained access to several models and could compare the answers to a real question. A model developer could learn which answer a user preferred. In its investment announcement, a16z argued that such feedback could complement fixed benchmarks and closed evaluation systems.[1]
Timeline
- 2024: Yupp closed a $33 million seed round led by a16z crypto, later disclosed when it emerged from stealth.[4]
- June 2025: The company launched its public comparison service.[2]
- March 2026: The founders announced the wind-down; users were given time to download their data.[2]
What They Built
A user entered a prompt and saw responses from multiple models side by side. The user could select a preferred answer and explain the choice. Yupp recorded those preferences, produced a public leaderboard and gave users credits for feedback. Its backers described digitally signed records intended to make contributions auditable.[1]
A reviewer for Tom's Guide tested the product in 2025 and described a scratch-card reward after feedback. Credits could fund further prompts, and the reviewer reported cash-out options. This first-hand account describes one period of the product; rewards and limits could change.[5]
Yupp's leaderboard explained that its score summarized user preferences and that model speed could affect voting. The distinction matters: a preferred answer is useful evidence, but a vote can reflect speed, style or the particular prompt as well as correctness.[6]
Market Position
Target Customers
Yupp needed two kinds of customer. Individuals needed a reason to compare models; AI labs needed data valuable enough to buy. The founders said several labs paid for the data, but they did not disclose customer concentration or revenue.[2]
Market Size
The sources reviewed do not establish Yupp's addressable market or how much labs spent on its specific type of feedback. The number of people trying AI tools is not a substitute for a count of organizations willing to buy preference data.
Competition
Arena, originally Chatbot Arena, also uses side-by-side comparisons and human preferences. Its research origins and leaderboard gave model makers another way to observe preferences. Yupp's distinct offer combined broad model access, rewards and a proposed data marketplace; the reviewed sources do not establish its relative data quality or cost.[7]
Business Model
Yupp intended to make anonymized preference data useful to model makers while giving individuals free access and rewards. Running multiple models and paying for feedback created costs; the company did not publish its gross margin, customer contracts or a breakdown of reward expenses. Several paying labs were not enough to demonstrate a durable business, according to the founders' shutdown account.[3] [2]
Traction
At closure, Yupp reported more than 1.3 million registered users and millions of preferences each month. Its legacy page lists over 30 million preference data points and 37 million prompts. These are company-reported totals, not independently audited counts or paid-customer metrics.[3] [8]
Post-Mortem
The two sides did not prove the same demand
The free comparison tool attracted a large user base. Its business depended on converting their feedback into data that labs would repeatedly purchase. Yupp did not disclose how much those customers paid or the cost of collecting their data, so the public record cannot identify the precise economic gap. The founders nevertheless said the offering had not reached sufficient product-market fit.[2]
The target moved beyond a single answer
Gupta said AI systems increasingly combine models with memory, tools and outside services. In his account, that shift reduced the importance of crowdsourced evaluation at the chatbot layer. It is a founder's assessment, not proof that preference data lost all value. Comparing complete workflows would require measuring actions and results as well as a single response.[2]
The contrast between Yupp's user count and its closure is instructive without implying those users were inactive. Large participation demonstrated interest in model access. It did not disclose whether that participation produced a repeatable, profitable product for labs.
Key Lessons
- Measure both sides. Yupp reported 1.3 million signups, but its data customers and their repeat purchases determined the business case.
- Preferences need context. Speed, prompt selection and answer style can influence a comparison; a leaderboard should explain what its score measures.
- Evaluate the workflow buyers need. The founder's account points to tools and memory as part of the system under evaluation, beyond an isolated chatbot response.