№ 100
Wednesday, September 2, 2026
AI & Tech Brief — September 2, 2026
№ 100
AI & Tech Brief — September 2, 2026
Anthropic launches Claude Fable 5.1 and Mythos 5.1 The same underlying model ships in two tiers: Fable 5.1 for everyone, Mythos 5.1 for vetted cyberdefenders and life scientists. Headline pricing is unchanged ($10/$50 per Mtok), but cache reads dropped 75% to $0.25/Mtok — Anthropic estimates ~25% savings on typical workloads and up to 45% on highly agentic ones. Fable 5.1 more than doubled Fable 5’s score on Terminal-Bench-Science (52.6% vs 24.7%), and it can now be used for defensive vulnerability discovery in source code (though pentesting and exploit generation still redirect to Opus 4.8). A new Enterprise Frontier Safeguards system lets enterprise customers store data on their own cloud infrastructure while Anthropic’s safety classifiers still run — effectively zero data retention with safety intact. The model also designed protein binders with ~50% hit rates (vs 10–15% typical) and mapped a third of Venus at 2–3 km resolution from 30-year-old Magellan radar data. Source: https://www.anthropic.com/claude-fable-and-mythos-5-1
OpenAI designates Astra as Critical cybersecurity capability — first model ever OpenAI’s “Path to Astra” post confirms the upcoming model meets the Critical threshold under its Preparedness Framework: it can identify and develop functional zero-day exploits across many hardened systems without human intervention. On an internal benchmark of 20 recent high-severity V8 vulnerabilities, Astra found and used two actual zero-days in exploit chains. Safeguards include 91.5% jailbreak refusal (up from 59% on GPT-5.6 Sol), activation classifiers, cross-conversation refusal training, and chain-of-thought monitoring for misaligned actions. Astra is the first model to undergo a formal U.S. government pre-release cybersecurity review. General capability ships normally; offensive cyber is gated behind Daybreak Blue with mandatory hardware security keys. Source: https://openai.com/index/path-to-astra/
Google launches agentic video understanding for Gemini Instead of ingesting video at a fixed 1 FPS, Gemini now decides what to watch, at what speed, and through which modality (frames, audio, transcript) — fetching only the segments relevant to the query. Available today on Gemini 3.7 Flash, 3.6 Flash, and 3.5 Flash-Lite via the Gemini API. Google reports up to 88% lower token consumption, 66% lower cost, and 7% better accuracy. The feature will also power YouTube’s “Ask YouTube” in the coming months. Source: https://blog.google/innovation-and-ai/models-and-research/gemini-models/introducing-agentic-video-in-gemini/
FBI probes service selling 153M+ driver’s license scans KrebsOnSecurity reports a dark-web service called Nexus is selling digital scans of over 153 million driver’s licenses from the US and Canada, apparently siphoned from identity verification company idscan.net (which serves Hertz, Target, FedEx, and 1,000+ marijuana dispensaries). Records include front/back scans with infrared and UV images, timestamped to when the license was last scanned — car rentals, dispensary visits, hotel check-ins. The FBI’s New Orleans field office opened an official investigation. The service vanished from the dark web hours after the story published. Source: https://krebsonsecurity.com/2026/09/fbi-probes-service-selling-153m-drivers-licenses/
Dan Luu reviews Ed Zitron’s AI skeptic predictions: wrong on everything Dan Luu systematically scored the most prominent AI skeptic’s predictions from 2024–2025. Every single falsifiable prediction was wrong: AI didn’t peak, OpenAI didn’t collapse, Cursor didn’t die (sold for $60B), CoreWeave didn’t go bankrupt, Meta/Google/Microsoft didn’t stop growing. Luu’s deeper point: the reasoning was as wrong as the predictions — numbers were thrown around for credibility but didn’t connect to coherent arguments. A useful case study in how gish gallop works and why confident, angry contrarianism gets rewarded with attention regardless of accuracy. Source: https://danluu.com/zitron/
Claude Code ships Fable 5.1 as default Fable model, plus a massive reliability update The September 1 Claude Code changelog is enormous: Fable 5.1 is now the default Fable model with 1M context and $0.25/Mtok cache reads. New features include a Containment Escape rule for auto mode, per-session effort overrides, time-format settings, and dozens of fixes for background sessions, subagents, sandboxing, and Remote Control. The VS Code extension got collapsible session management and a model picker in the input footer. Source: https://code.claude.com/docs/en/changelog