Rosetta Intel

OpenAI / Anthropic / Google DeepMind product and model news; security-relevant items marked 🛡

OpenAI

No major official releases captured in this window. Industry tracking shows roughly 9 new releases across the three labs in July (including frontier models, coding, open source, and agents), with no confirmed same-day items on the OpenAI side.
ThursdAI July 2026

Claude Fable 5 restored globally: The U.S. Commerce Department lifted its export controls on Fable 5 on 6/30; from 7/1 it is fully back on Claude.ai, Claude Platform, Claude Code, and Claude Cowork. The relaunch ships with deeper cybersecurity protections, a draft AI jailbreak-severity scoring framework, and a HackerOne program for security researchers to report "cyber jailbreaks."
The Hacker News
Jailbreak-severity framework: Anthropic, together with Glasswing partners including Amazon, Microsoft, and Google, proposes an industry-wide jailbreak-severity scoring standard, accompanied by a new classifier with a >99% block rate against specific jailbreak techniques.
Anthropic Newsroom
Claude Sonnet 5 launched: Performance approaching Opus 4.8, with $2/$10 per million tokens introductory pricing (through 8/31). Safety evaluations report a lower rate of undesirable behavior than Sonnet 4.6, safer agent scenarios, and cyberattack capability significantly below the current Opus.
Introducing Claude Sonnet 5
Intellectual-property dispute: Anthropic alleges that some Chinese companies extracted knowledge from Claude via "distillation" (a model-security / IP angle).
Washington Post

NanoBanana 2 Lite: Image generation from <4 seconds, $0.034 per 1,000 images, with quality above the original NanoBanana; the Interactions API went GA the same week.
ThursdAI July 2026
OmniFlash: The first of Google's any-to-any Omni family, generating video up to 10 seconds and supporting multi-turn conversational refinement (via the Interactions API).
ThursdAI July 2026

🛡 = security-relevant