AI Prompt
BlogAbout
← Back

Claude Code's Product Lead Joins OpenAI, Pierre, and More

4 min read

The story that stood out on Thursday was a people move. Cat Wu, who led product for Claude Code, has left Anthropic and now leads product management for OpenAI's Core Products, working with Tibo Sottiaux. Around the same time, Austin Lau, who built Anthropic's growth function, announced that he had also joined OpenAI.

I don't read too much into any single hire, but when the person who shaped the product I use every day for coding changes sides, it's worth noticing. I'll be watching what the Codex product looks like over the next few months. (Cat's announcement, Austin's post)

What Else Happened

OpenAI will acquire Pierre Computer Company to deepen Codex. Pierre makes Code Storage, Trees, and Diffs, and those libraries already power browsing and review inside Codex. Code Storage is headed into the product after the deal closes, and Pierre's open-source work is staying open, which is the part I'd hold them to. (OpenAI announcement)

Artificial Analysis replaced tokens per second with Time per Task. The independent API benchmark now ranks about 1,000 serverless endpoints by how long a model takes to finish a job instead of how fast it streams. GPT-6 Astra at max effort finishes in about 17 minutes, versus about 26 for Opus 5.5 and about 29 for Gemini 4 Carbon, despite slower tokens per second, because it writes far fewer tokens. Default medium effort cuts time and cost several times over with a modest drop in intelligence. This is a much better match for how agents actually get used than raw streaming speed. (Artificial Analysis)

Figma beat Q3 and raised its full-year guidance. Revenue hit $391.4M, up 43% year over year, with net dollar retention at 131%, and 2026 guidance moved up to about $1.497B to $1.499B. Dylan Field's line was that as code gets cheaper, design becomes the scarce layer. MCP is at roughly 1M weekly actives and already drives about 5% of AI credit use. (Figma on X, CNBC)

ChatGPT Images 2.6 is out. It adds GIFs, reusable reference images, and finer edits, and GPT-Image-2.6 Flare and Sunburst reached the API with about 50% faster generation. It's live in ChatGPT and Codex across tiers. (OpenAI)

Ricursive Intelligence open-weighted Ripple-65B Preview. It's a 65B dense model trained from scratch on 12T tokens for about $32M. The lab claims U.S. open-weight state of the art on agentic coding, at 80.5% on Terminal-Bench 2.1, while fitting on a MacBook or DGX Spark. Those are the lab's own numbers, so I'd want independent runs before leaning on them. (Ricursive, ricursive.com/ripple)

Together AI introduced Together Compiler. It does continual "polymorphic" inference optimization, so overnight agent runs can match what would otherwise take weeks of manual tuning as models and traffic shift. (Together AI)

Oumi launched FAiE, a Frontier AI Engineer. It builds, evaluates, trains, and improves custom agents using SFT, DPO, and a GRPO preview, and it supports Bring Your Own Harness. (Oumi Fall Launch, Business Wire)

What I'm Reading

  • WordPress Studio redesign – a new workspace with an agent and live preview, bring your own agent, MCP for Claude, Codex, and Cursor, annotations, and a DESIGN.md; free and open source on Mac, Windows, and Linux.
  • Coincide – a native Swift and AppKit Mac app that puts Claude Code, Codex, and Cursor side by side and lets you switch mid-thread; free and open source, at coincide.dev.

Accounts Worth Following

  • Rich Tabor (@richtabor) – shipping WordPress Studio as a design-first, agent-ready local workspace, with clear demos of where product UI and coding agents meet.
  • Apoorv (@apoorvdarshan) – building Coincide, a polished native Mac shell for multi-agent coding, with short demos and open-source shipping notes.
  • Ricursive Intelligence (@RicursiveAI) – the small lab behind Ripple-65B, with tight posts on efficient open coding models and real benchmark numbers.

That's it for today. 👋