How AI benchmarks actually work — and why a 0.46-point lead is not a verdict
Vals, SWE-bench, ARC-AGI-2, METR, GDPval, HELM and LMArena answer different questions. Here is how to read a leaderboard without mistaking a score for a capability.
Artificial intelligence is the beat SoilTune follows closest: the models, the money, the hardware and the rules that decide who gets to build with it. This page collects every AI story we publish, newest first. New to the desk? Start with the reading path below.
Start here
The price of a million tokens — how API pricing split in 2026, and what a support workload actually costs.
How AI export controls actually work — why the chip is no longer the border, and how TPP, D:5 and data-centre end uses decide who can buy what.
Claude picked a bombed Iranian school as a target — what a 32-run test found about models used for targeting.
Apple is building a computer-use agent — the job listing that describes how it works.
Google puts your AI chat history on the homepage — the quiet change to AI Mode.
Baseten’s bison-1 tops the decision-model index — a smaller lab winning on structured decisions.
Dutch regulator tells shoppers not to buy AI camera glasses — Europe’s first consumer warning on AI wearables.
Google opens Gemini 4 Carbon to all paying developers — the discount and its deadline.
Banks bar Chinese and Hong Kong investors from Anthropic’s IPO — where export controls meet capital.
The RTX PRO 5500 runs the same GB202 configuration as the RTX 5090 — the hardware layer under the models.
We report what is documented, link the primary source, and label what is still unknown. No benchmarks we have not run, no forecasts dressed as facts.
9 stories
Vals, SWE-bench, ARC-AGI-2, METR, GDPval, HELM and LMArena answer different questions. Here is how to read a leaderboard without mistaking a score for a capability.
California’s official layoff filings cover 31,148 workers in five months. Tech is a minority of them,…
A standard mix of input and output now costs $0.10 to $16 per million tokens. The…
Two open-weight models built on Gemma 4 return calibrated probabilities instead of prose. bison-1 now leads…
After testing six pairs, the Autoriteit Persoonsgegevens found dangerous advice, hidden recording lights and cameras strangers…
A collapsible sidebar now lists recent AI Mode threads beside the search box. Google says it…
A senior engineering manager posting described agents that browse, run commands and write code. Apple pulled…
Vals AI tested 13 models on real satellite imagery of a bombed Iranian school. Claude Opus…
Gemini 4 Carbon is open to paying developers at $2/$10 per million tokens until Nov 30.…