Horizon IA - The AI news you need

A hundred DeepMind agents cleared 34 fake proofs in 27 minutes, then a quarter of the swarm documented the fraud with no instruction.

Our GPT-6 Astra test: the record score comes from a custom harness, and caching decides the invoice far more than the ranking does.

Our Claude Fable 5.1 test: headline rate unchanged, cache reads quartered, and three API breaks to plan for before migrating.

GPT-6 Astra ships at 10 dollars per million input tokens, with a 99.9% ARC-AGI-3 score that drops to 62.7% on the standard harness.

Astra's reasoning now runs partly inside an unreadable internal loop, and OpenAI rates the model critical for cybersecurity.

Gemini 3.8 Flash scores 73.7% on DeepSWE v1.1, three tenths behind Claude Opus 5, while its Cyber variant reaches 86.2% on CyberGym.

Claude text detection enters private preview: regulators, newsrooms and fact-checkers can now request the key that reveals the watermark.

Claude Fable 5.1 climbs from 24.7% to 52.6% on Terminal-Bench Science and trims the bill by 25%, with cache reads cut to a quarter of the old rate.

95% of China's 128,000 first-quarter short dramas are AI video. A generated minute costs 90% less than shooting the same scene with actors.

Claude Code limits rise 25% on September 14, but the summer's 50% boost ends the same day. Anthropic puts the net loss at 17%.