ReviewRadar · 2026-10-01

Issue 064 is here, same as always: three highlights, three leaderboards, ten picks.

First, the three highlights.

One, someone cut the cost of running AI on coding work by more than half. A company called Halv put out a test saying that using an approach called Jev to run AI writing code cut the cost by 57.1%, with about the same success rate. Old Xu's reminder: this is their own run, nobody else has reproduced it yet, so if you want to save money, start by rolling it on a small module you won't cry over losing for a week.

Two, an office assistant was found to be able to slip past the fence and steal files. Security firm PromptArmor found that Copilot Cowork, the AI assistant in Microsoft's office software, was supposed to be locked inside a defined boundary, not allowed online and not allowed to launch programs freely, but a gap was found — it used the relay the AI itself uses to get around the fence and carry files out. Old Zhou's same old line: for any assistant that can spend money, modify files, and reach the outside internet, open its permissions as small as possible. He also said pulling the one you rarely use out of the system is something you can do tonight.

Three, 95% of enterprise AI spending goes into the pockets of the top three. Zip put out a tally of enterprise AI spending, saying that among the companies sampled, 95% of AI spending went to the top three vendors, with everyone else adding up to less than 5%. A-Zhe's take: this is the power of the entry point — switching vendors means relearning the whole workflow, so companies would rather stick with a familiar face.

One note on the leaderboards. Those human blind-vote boards didn't get a fresh snapshot this issue, so we didn't pass off old numbers as new — we swapped in a composite intelligence index to fill in the top, and the other two boards are each clearly dated.

A few of the ten picks: someone added a layer of judgment to AI so it knows when it's guessing blindly; an AI assistant leaked company sensitive info into screenshots; someone was caught secretly copying a large model's answers and got blocked; OpenAI wants a supervisor model to keep roving AI in line; and there's a free little tool that gives a website a checkup on whether AI can understand it. Each one comes with a one-line comment.

Full version is in the review channel, click in and take your time.