AI Radar

This is the first time delegation is genuinely asynchronous. Not "the agent works while I watch", but "the job is done by morning". Anthropic’s own usage data shows where it lands: 33.4% business process and operations, 16.4% content and copywriting. That is exactly the work an agency still does by hand.

Give an agent a folder, never your drive. Anthropic itself warns about destructive actions on ambiguous instructions and about indirect prompt injection through files and websites. Human review on anything outbound — mail, publishing, payment — is not caution, it is design.

The mini and nano suffixes are gone. Sol, Terra and Luna are durable capability tiers that can each advance at their own pace. In practice: you route per task on price rather than per project on model name, and your prompts survive the next release.

Put Luna on classification, extraction and bulk summarisation; Terra on editorial work; Sol only where reasoning is genuinely worth the money. The gap between Luna and Sol is five times on input and five times on output.

"Open" is not the same as "self-hostable". Even at four-bit precision K3 needs about 1.4 terabytes of fast memory — that means Blackwell or MI400 server nodes, not a workstation under your desk. The bill shifts from API subscription to hardware and operations, and for most agencies that is the worse trade.

Treat open weights as leverage, not as a migration plan. They push hosted prices down and give you an exit under vendor risk. You run them at a provider that already owns the nodes.

For an agency, picture-plus-sound in one generation removes an entire edit step. No separate voice-over laid over the footage, no lip-sync correction afterwards. That an image lab is simultaneously moving into robotics says something about where generative models are heading: from picture to action.

The evaluations are early — the lab says so itself. Use FLUX 3 for pitch material and social where twenty seconds is enough; keep Veo or Kling alongside while you wait on gated access.

OSWorld measures computer use: clicking, forms, real screens. That is the benchmark that predicts whether an agent can take over your work, far more than a knowledge quiz. A jump there at a third of the cost changes which tasks you are willing to delegate at all.

Opus 5 is the new default on Claude Max and the strongest model on Pro. If your own jobs still run on a 2025 model you are paying the same for less. Re-benchmark your two heaviest workflows this week.

The gap between "half your market searches here" and "14% measure it" is the whole story. Whoever can answer today, with numbers, what share of demand runs through AI answers is selling a measurement, not a trend.

Start with a baseline: for which questions are you cited in ChatGPT, Perplexity, Gemini and Google’s AI overviews, and from which source? That is a report you can produce today and almost nobody has.