The big model news
Google shipped Gemini 4 Carbon on October 28 — its strongest model yet, built specifically for cybersecurity and long-horizon engineering work. It leads the field on SEC-bench Pro (88.2%) for finding vulnerabilities, ExploitGym (54.8%) for exploitation, and CWE-bench (71.0%) for patching, and Google says it tops the latest Artificial Analysis Intelligence Index at a score of 59. The catch: it's not available to the public yet. Access starts with a small group of cyber defenders and widens gradually through Google's Fairwind Program, with no date set for developers or consumers. Pricing when it does arrive: $2/$10 per million input/output tokens introductory, rising to $4/$20 later.
Mistral released the open weights for Mistral Large 4 ("Le Chonk") on October 29, a 1-trillion-parameter multimodal mixture-of-experts model with 49B active parameters and a 1M-token context window. Mistral calls it the best open-weight coding model available, trained entirely on European infrastructure. It ships under the Mistral Ecosystem Open-Weight License — free to use, modify, and commercialize, with a carve-out requiring a separate license if you're running it as a paid model-as-a-service. API pricing is $0.70 input / $0.07 cached / $2.10 output per million tokens, and the weights are downloadable on Hugging Face today.
Anthropic's IPO paperwork went public on October 26. The S-1 shows $16.3B in first-half revenue against an $80.7B net loss, and reveals a governance structure built to keep founder control after listing under ticker ANT on Nasdaq. No shares trade yet — the filing just makes the numbers real.
Product Hunt
The week's standout launch was Dedica, which hit #1 Product of the Day on October 28 with 307 upvotes. Its pitch: AI agents like browser assistants now shop, sign up, and fill out forms on people's behalf, and most sites silently fail them. Dedica runs 125 automated checks against your site the way an agent would use it, scores it out of 100, and hands you a fix-it file to drop straight into Claude Code or Cursor. Free to scan, $30 for the fix kit.
Two days later, Vaani Developer Platform took the top spot (388 points) — a no-code harness for building voice agents that handle real phone calls, manage objections, and hand back structured data when the call ends. Also worth a look from that batch: Chopstick, which watches what you repeat on your Mac and quietly builds notch widgets for it using whatever coding agent you already have.
GitHub trending
Developer-tooling-for-agents is the theme this week. claude-mem (thedotmack) is climbing fast — it captures everything your coding agent does in a session, compresses it, and re-injects relevant context into future sessions, with support across Claude Code, Codex, Gemini, and more. hallmark (Nutlope) is an "anti-AI-slop" design skill for Claude Code, Cursor, and Codex aimed at the generic-looking-output problem every vibe coder has hit. And Scrapling keeps climbing as the go-to scraping library for agents that need to pull structured data out of messy pages.
Worth your time
If you're deciding where to spend attention this week: Mistral Large 4's open weights are the most immediately useful thing here — downloadable today, genuinely strong at coding, and self-hostable if you care about data residency. Gemini 4 Carbon is worth watching but not worth waiting on; it's gated and there's no public timeline. And if you ship anything agent-facing, Dedica's 125-point audit is a five-minute gut check worth running against your own site before an AI shopper bounces off it.