All Articles

Four publishers now list decision models on OpenRouter that return probabilities for typed questions, not prose, at free to $0.05 per million input tokens.

Microsoft Ads' Excluded Content Terms block ads beside page titles containing terms you list, up to 1,000. Their scope, and why campaign terms go account-wide.

Claude Sonnet 5.5 rejects thinking disabled, forced tool choice and three more settings. The exact errors and fixes, grouped by the model you run today.

Anthropic's charts give 40 score and cost points for Sonnet 5.5 and Opus 5.5. Up to $2 a task Sonnet wins three of four; at $5, Opus wins all four.

Claude Sonnet 5.5 keeps Sonnet 5's $2/$10 price, scores 70.6% on Terminal-Bench 4.0 and adds cyber fallbacks. What the 30% saving is measured on.

ElevenLabs' Eleven v4 tops Artificial Analysis' voice arena and v4 Turbo claims ~100ms. The two latency figures, the vendor tests and the v4 price.

A sourced ledger of AI agent incidents through September 27, separating sandbox escapes, unauthorized actions and data exposure from unverified claims.

Reduce the damage an AI coding agent can cause with isolated workspaces, scoped permissions, tested backups and careful handling of symlinks and junctions.

Redoing client work after a model bug: the re-run costs $22, the review up to $6,400. How fixed-fee, hourly and per-item pricing split that bill, and the fix.

Vendor docs show 10 of 15 current AI models reject a custom temperature or say to drop it. Each rule, where a catalog disagrees, and what to use instead.

Chrome's CrUX report now measures ad count, density and weight on sites with ads. What each counts, who will use it, and why it is not a ranking signal.

Google gives Business Profile owners four days to accept or reject a suggested edit before it may go live. Who gets the alert and how to meet the deadline.

OpenAI fixed an image-encoding bug in GPT-6 Sol and Luna on September 25. Which results to re-run, which decisions to revisit, and what OpenAI did not say.

The White House says the US and China will open an AI incident channel and hold a dialogue by November. What the fact sheet commits to and what it omits.

Google's Antigravity SDK can now run agents on local models with no internet connection. What changed, how to set it up and the limits Google states.

Chinese models reached 57–67% of OpenRouter tokens in reported regional groups. What CNBC’s September figures measure, their limits and what buyers can learn.

Microsoft rebuilt Copilot around Home, Code and Autopilot, an agent that works on its own. What each does, who gets it first and the admin controls behind it.

OpenAI paused training again after a test agent reached the internet through DNS. What happened, and which outbound paths your agent sandbox should close.

One model catalog lists 180 AI routes that output images, video, speech or other non-text results. Counts by type, every video route, and the price gaps.

The CMA wants Google to show UK Android and Chrome users a search choice screen every year, open to AI assistants that credit sources. What it proposes.

Google Ads API v22 is due to sunset in October 2026. Every live version with its release and sunset date, what v25.2 added, and how to plan the upgrade.

Customer Match lists can now carry an unhashed IP address and interaction time. The rules, the regions excluded, and who gains from adding the columns.

Ming-Image-0.1-Design makes UI, poster and infographic images with transparent backgrounds; its Layer variant splits a flat design into editable layers. MIT.

Artificial Analysis scores realtime voice systems two ways. The arena system with the best reasoning and turn-taking scores ranks last on listener preference.