Yesterday, Anthropic quietly released Claude Opus 4.7 — their most capable publicly available model to date. It's sharper at complex coding, better at understanding images, and more creative when drafting documents and slides. Early testers include Cursor, Replit, Notion, Shopify, and Intuit. But here's the twist: Anthropic openly states that Opus 4.7 does NOT push their capability frontier. Their real frontier model — Claude Mythos Preview — is more powerful on every benchmark. It's so advanced that Anthropic is keeping it locked away, available only to a handful of trusted partners like Apple, Google, Microsoft, Nvidia, and JPMorgan Chase. Why release a less capable model first? Deliberately. Anthropic is using Opus 4.7 as a testbed for new cybersecurity safeguards before they dare open Mythos-class capabilities to the public. Think of it as a controlled detonation — better to stress-test the safety measures on a slightly smaller bomb. For security researchers who want access to the full power, Anthropic launched a Cyber Verification Program — essentially a vetting system to lift some of the restrictions. The AI race isn't just about who builds the most powerful model anymore. It's about who can prove they can deploy it safely enough to actually use it. Anthropic just drew a very clear line in the sand.

Researchers found that Claude Opus — a model anyone can access right now — successfully wrote a working Chrome browser exploit. Cost: $2,283 in API calls. This comes as Anthropic separately withheld its Mythos security model from public release, citing danger. But if the publicly available version is already doing this, the line between "safe" and "dangerous" AI is a lot blurrier than we're being told. ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━ Source: https://go.theregister.com/feed/www.theregister.com/2026/04/17/claude_opus_wrote_chrome_exploit

Custom hooks: scripts that fire on system events — messages, resets, compaction. The agent doesn't know they exist. It can't bypass them. Advanced 5 is live.

GPT-Rosalind is a frontier reasoning model built for life sciences — genomics, protein analysis, drug discovery workflows. It's OpenAI's clearest signal yet that specialized vertical AI is the next frontier, not just bigger general models. If this works even half as well as claimed, pharmaceutical R&D timelines could compress dramatically. https://openai.com/index/introducing-gpt-rosalind

Real RAG: PDF → vectors → semantic search. Your agent doesn't read documents — it searches a knowledge base. $0.01 to index, $0.0001 per query. Advanced 4 is live.