Coding-related use-cases are only 25% of the methodology of their updated Intelligence Index.

Gap between TerminalBench and FrontierCode is basically a "read my mind and correctly guess what my underspecified 7-word prompt means" benchmark. No surprise that Opus leads in that; Anthropic has geared their model more toward understanding people over systems for a while.

Just a bunch of <= 7-word underspecified "read my mind please" types of prompts.

Seems like a theme.

Still cooking on a DeepSeek V4.1 Flash recipe for 4x DGX Sparks. Pretty significant wins on single-stream decode compared to competitors, and absolutely dominates in prefill.

People who start fights and get themselves all upset enough to block you... lol. I don't understand why these people even bother starting arguments and upsetting themselves so much. Created their own problem.

Most of the open-weight losses come at the hands of frontier STEM weakness rather than everyday coding and similar work. It makes sense. They're more like 110 IQ office workers rather than 140+ IQ PhD rocket scientists.

This is the methodology btw.

Why does every dude have 3-5 different phones lol.

2. "Don't trust the client" has been a thing in SaaS for a LONG time. Without even trying to dig, a simple search for "don't trust the client" surfaces a top result written as early as 2015. If I really wanted to find the first utterance, I could go WAY back.

OK, so a couple of things: 1. You're overstating the problem. This is a problem for you, and you SPECIFICALLY, because your target customer is B2B, technical, and is (inherently, by the definition of your product) using AI heavily.