Research
Method, numbers and code, measured against a sealed key.
9 posts
ResearchAI trading bots are a lie, and you should know about it
We checked 30 sellers of AI trading bots, read 21 more from a second report, read the regulator files, and ran five tests of our own on Polymarket weather bets, the sold recipes, an overfitting demo, the S&P 500 and Bitcoin. No public, audited, after fee record exists. Timestral, a time series AI, did not beat the simplest bet once in 16 tries.
Nexus Agent TeamOct 21, 2026 · 38 min
Listen · 18 minResearchMistral's new model predicted hotel room prices a year ahead with 88.6% accuracy
Hotel pricing teams still check prices by hand: 10 hotels per person, 2 hours a day. We tested Mistral's new Timestral model against six other tools on public hotel prices. A year ahead it was the most accurate, at 88.6%, and its price ranges held. On 175 Spanish markets it beat the classic tools on 160 to 165. What we found, explained from zero, and what a hotel group can do on Monday.
Nexus Agent TeamOct 20, 2026 · 17 min
ResearchHaiku 5.5 matches Sonnet 5.5 on retrieval. It costs a fifth as much.
We ran Haiku 4.5, Haiku 5.5 and Sonnet 5.5 at every effort level on 18 real retrieval questions and 10 real filing jobs, scored by a blind judge. Haiku 5.5 at high effort scores 71.7% on retrieval at $1.15 per attempt. Sonnet 5.5 at high effort scores 71.3% at $5.57. We are moving our retrieval and filing agents to Haiku 5.5.
Nexus Agent TeamOct 9, 2026 · 24 min
ResearchOpenAI launched a Jev competitor. It is slower, costs 2× more and is no more accurate.
OpenAI launched the Decisions API, a competitor to TypeSafe's Jev. We tested both on the same cases: three timed runs and a token pass, more than 9,000 calls. The Decisions API is slower at the tail, costs 2× more per call and is no more accurate. We also list the rivals that now sit beside them.
Nexus Agent TeamOct 7, 2026 · 25 min
Listen · 29 minResearchNearly 1 in 4 tests Claude wrote were useless. Then we made our test runs 9x faster
Nearly 1 in 4: in two apps built by Claude coding agents, 23% of the browser tests could not fail on the product's own features: 133 of 578 tested a demo the product does not have, or were empty. Across three builds, the tests caught 0 of 65 bugs before a human did. Then we made our own test runs about 9x faster, from 55 minutes to about 6. What went wrong, explained from zero, and what to change on Monday.
Nexus TeamSep 30, 2026 · 22 min
Listen · 22 minResearchThe guard that works while the agent works
A judge that reads meaning, next to code checks that read facts, watching an agent while it works for days. We ran it live on long builds. Most of what went wrong was not the model. This note gives the three moments, the contract every judge needs, and the six pitfalls that cost us the most.
Nexus Agent TeamSep 26, 2026 · 21 min
Listen · 24 minResearchJev: a boring classifier that unlocks a lot
Jev is a classifier: it picks an answer from a list. That sounds boring. But it is fast, cheap and accurate at the same time, and that opens up new uses. Here is how it works, what people built with it in two days, and what happened when we used it to check our own AI assistant.
Nexus Agent TeamSep 17, 2026 · 26 min
Listen · 20 minResearchThe write side of memory
Your company's durable asset is what it knows, and most of that is written down nowhere. AI agents can now do the writing: file what work teaches, continuously, at scale. Should you trust them with your company's memory? We invented a company, planted 24 traps in its paperwork, set four filing agents loose on it at once, and hired an adversary to break what they shared. Part 2 of our agent-memory research: what held, what silently vanished, and the four ways a rulebook fails.
Nexus Agent TeamJul 6, 2026 · 43 min
Listen · 18 minResearchOne judge, many readers
How should AI agents read a company's files, and how should they write down what a company learns? We built a realistic knowledge drive, seeded it with seven traps, and ran eight agent architectures against it. Five ways to read. Three ways to file. Every run measured for speed, computation, and quality against a sealed answer key. This note explains every concept from zero and ends with an architecture you can adopt.
Nexus Agent TeamJul 5, 2026 · 31 min
Start with one procedure
Tell us where the work piles up.
A Nexus engineer builds the first procedure with your team and shows it running in your own tools. Four weeks to a procedure that runs itself. And a number you can defend.
Talk to our team