Ephemeris.
Today intelligence gets a price tag, a usage index, a behavioral score, and a map of what it can say out loud.
GPT-5.6 shipped July 9 as a family, not a model: Sol for hard coding and agents, Terra for the middle, Luna tuned for latency. Sol tops the independent Coding Agent Index at 80 and hits 88.8% on Terminal-Bench 2.1 — above GPT-5.5's 83.4% — and it's already the default across Codex, the API, and Copilot.
Sol Pro and Sol Ultra add heavier compute at the same model and price. 5.6 is the last of the 5.x line before GPT-6.
Google put two media models in the Gemini API. Omni Flash makes 10-second clips at $0.10 a second, taking text, images, and up to three seconds of reference video; Nano Banana 2 Lite returns a 1K image in about four seconds for $0.034 per thousand. Both are reachable from AI Studio, Flow, or a single call.
Vercel's AI Gateway sees real production traffic, and July's index reads as a map of demand. Open-weight models jumped to 29% of tokens from 11% in April — while taking under 4% of the money. Anthropic runs 32% of tokens but 61% of spend, and 72% or more in every high-stakes use case.
Precursor injects an obfuscated script that watches a whole session — pointer paths, keystroke rhythm, focus, visibility — instead of a single checkpoint. It scores the human tells bots can't fake at scale: wrist-pivot geometry, cognitive-load delays, the micro-tremor of a real hand. A page refresh no longer resets the signature. Keystrokes are read as timing, never the keys themselves.
A new Anthropic paper maps "J-space" — the concepts a model can actually verbalize — with a Jacobian Lens that measures which internal features steer the output. Zvi's read: about 25 concepts sit there at once, but only ~6 are truly distinct. Swap one ("France" → "China") and its capital, language, and currency shift with it. Automatic, System-1 answers skip the space entirely.
Five picks from the morning feeds, read so you don't have to — a new model family, cheap generative media, the market's real usage, a behavioral bot check, and a look inside the model's own reasoning.
Sources · OpenAI · Google (via The Batch) · Vercel · Cloudflare · Don't Worry About the Vase
Rubric · tools to adopt · generative media · dev & market signal · privacy & security · research with a working edge
Issue · 085 · 14 July 2026 · Zürich