Launch offerSix months of the Starter plan, free for teams who sign up by November 22Claim it →

Blog

Notes on how AI agents fail, what actually improves quality, latency, and cost, and the teams shipping on evidence instead of vibes.

Cover image for three.dev is now generally available
Product

three.dev is now generally available

three.dev is now generally available. It automatically finds how your AI agent fails in your live traffic and hands the fix to your coding agent.

Bryan LeeBryan Lee ·
Cover image for Your use case is not in the benchmarks
Product

Your use case is not in the benchmarks

We replayed 2,000 requests from a use case no benchmark covers and measured how seven frontier models performed across quality, latency, and cost. The winner is one no benchmark would have picked.

Andrea MoscatelliAndrea Moscatelli ·
Cover image for The free upgrade that raised your bill
Engineering

The free upgrade that raised your bill

Four hidden factors can move your bill as much as the price per token, or more. On our 191 requests, a 33% price cut came back as a 4% bigger invoice.

Dani Rodríguez HernándezDani Rodríguez Hernández ·
Cover image for Detect and fix your agent failures
Product

Detect and fix your agent failures

three.dev automatically discovers how your AI agent fails, then gives your coding agent what it needs to remediate those failures properly. In our case study, the result was a fix more than three times as effective as the standard process.

Andrea MoscatelliAndrea Moscatelli ·
Cover image for Jev vs the AI Judge: can a decision model replace it in production?
Engineering

Jev vs the AI Judge: can a decision model replace it in production?

TypeSafe's Jev answers questions about a text without writing a word. We tested it against three.dev's scoring on two evaluation scenarios and roughly 15,000 requests: can it replace that scoring, and can it say what went wrong?

Raúl Moreno SalinasRaúl Moreno Salinas ·
Cover image for Knowledge just moved out of the GPU
Engineering

Knowledge just moved out of the GPU

DeepSeek and Qwen just agreed on where a model's facts should live: in host RAM or on an SSD, not in GPU memory. Here's how that split changes what it costs to run a model, and what it means if you want to own a specialized one.

Ximo GuanterXimo Guanter ·
Cover image for Your business can fit in a smaller model
Engineering

Your business can fit in a smaller model

An introduction to the world of open-weight models for businesses wondering whether they could own their AI, where a specialized model trained on your own data can match the big ones at a fraction of the cost and latency, and where the data you already have matters more than any model you could pick.

Dani Rodríguez HernándezDani Rodríguez Hernández ·
Cover image for Hello, world. We're round3.
Company

Hello, world. We're round3.

We started a company to answer the hardest question in production AI: is this change actually better?

Borja BurgosBorja Burgos ·