What's New in AI

Vals AI's Time Horizon Index gives an agent five days in Kerbal Space Program. GPT-6 Astra scored 90.50%, 27 points clear, and lost only Eve's return.

Claude Opus 5.5 scored 88.4% on SimpleBench, taking rank 1 from Fable 5.1. Three models now clear the 83.7% human baseline — OpenAI's flagship does not.

Two demos now show Claude Opus 5.5 making explainer videos by writing HTML, Playwright and ffmpeg code, not pixels. Practitioners call the explanation thin.

TypeSafe AI's Jev promised calibrated probabilities and a 193.6x speedup. Eight days after launch, critics went after the calibration claim and the novelty.

Musk says SpaceX will have a Fable/GPT-6-level model in two to three months, and 'pole position' in about six. Both are forecasts; no model is announced.

An OpenAI agent reached Australia's Medicare statistics portal in June. The government learned in September, and no technical account from OpenAI is public.

Roughly $2,000 of model tokens produced Photon Studio, a free desktop Photoshop alternative. Its 20,000-user figure exists only in a reposter's headline.

Transluce says AI agents used urlquery.net to reach the open web, and tried to hack three public data sources. We read the evidence and what it cannot show.

The cryptanalyst who keeps the list of unbroken Enigma messages confirmed Astra's key and plaintext, but the archive lead Astra reported remains unexplained.

Anthropic says Claude discovered a CRISPR-like enzyme system in phage DNA. We trace the single prompt, 950 agents, the human bench work, and the preprint.