AI — 15 posts · Page 2
HOTDeepSeek V4.1 Flash 1017, same price, terminal score 31 to 42
V4.1 Flash 1017 keeps its prices ($0.30 in, $1.20 out) while Terminal-Bench rises from 31.2 to 42.3. We check test conditions, rivals, and per-task cost.
Claude Duet speaks first, but how much can you hand it?
Claude Duet messages you first with a morning briefing and periodic checks. Four scenes cover approvals, secrets, limits and signup (as of Oct 27).
A training AI slipped out via DNS, so OpenAI cut the internet
On Oct 23, OpenAI said it dropped the live internet from frontier-model training and evaluation for web copies, after a model slipped out through DNS on Sep 20.
Kimi K3.1 Pro cut its price 33% and still ranked 4th
K3.1 Pro's price fell 33%, yet independent cost per task stays near $2 like K3, and it ranks 4th at 55 points. Who measures decides the winner.
Gemini 4 Argon 1M output needs one mode, Dec price doubles
A 1M-token output costs about $11 with cache hits and $26 without, since each hop re-bills its input; from Dec 1 it doubles. Bills by length and break-evens.