Claude's Ultra Plan, Vercel's Agent Desktop, and More
The change most of us will feel is a pricing one. Anthropic renamed its Max plans: Max 5x is now just Max, and Max 20x is now Ultra. Prices stay the same at $100 and $200 a month, and existing subscribers don't need to do anything.
The bigger part is that Anthropic is removing the 5-hour limit for verified Ultra accounts, starting with most existing subscribers. Newer accounts keep a temporary limit until they're verified, and weekly limits don't change. The updated pricing page also lists monthly API credits on both plans, $100 on Max and $200 on Ultra. If you run long Claude Code sessions, the rolling 5-hour window was usually the wall you hit first, so this matters more than the rename. Claude on X, pricing page.
What Else Happened
Vercel Labs open-sourced agent-desktop, computer use for any coding agent. Chris Tate pitches it as "agent-browser, for your whole computer." It's a Rust CLI and C library for macOS, Windows, and Linux that reads each app's accessibility tree instead of guessing from pixels, so every element gets a stable ref the agent can act on and then re-check. It runs in the background while you keep working, and it ships its own agent skills inside the binary. (Chris Tate's launch, GitHub repo)
Anthropic set thousands of Claude agents loose on its own training-data monitor. In a deliberately one-sided game, Claude had to write fine-tuning data that looked harmless to a production-style monitor but made a small research model misaligned. Given enough rounds, it succeeded, mostly by hiding the behavior in fiction and roleplay, like a beekeeping Q&A or a veteran wrestler mentoring a newcomer. Anthropic is clear these models were trained to misbehave only for the experiment. The point is that safety testing can scale with compute instead of researcher hours. (Anthropic on X, Alignment Science post)
OpenAI shipped a batch of Agents API updates. Auto Review checks commands for bad behavior before they run, like it already does in Codex. Async Compaction moves context compaction into the background. An experimental time-awareness feature gives the model a sense of time passing, so "timebox CI fixes to 10 minutes" actually means something. You can also swap a session's tools mid-session, and system prompts are now cached across sessions automatically. (OpenAI Devs, Steven Coffey's rundown)
OpenAI also launched Private Intelligence for businesses. It bundles stricter limits on OpenAI's access to business content, Zero Data Retention, and a Private Inference preview built on confidential computing. It also lets companies set which data agents can reach, which actions they can take, and when they need approval. It's aimed squarely at the "we can't send that to a model" objection. (OpenAI on X, Private Intelligence page)
Google's latest Gemini Drops bring skills and Antigravity into the Gemini app. Skills let you save instructions once, build them straight from a chat, and stack them for bigger tasks. The Mac app can now jump from a chat into a local Antigravity coding environment, and PDFs, Docs, and Sheets can be edited right in the conversation. (Gemini app thread, skills, Antigravity on Mac)
Lauren Tan argued that code now arrives as a stream, not a drip. Her point is that agents give every engineer an elastic workforce, so teams have to design systems where agents with little context still do the right thing: "make the easy thing the right thing." It pairs well with John Carmack's note that good building blocks matter, and that agents should be trained to keep them clean, not just to finish tasks. (Lauren Tan's post, Carmack's post)
Launches and Demos
- Geist Serif – Vercel's new high-contrast display serif, with the same cap height and x-height as Geist Sans. A variable axis swaps every dot in the font between round, square, triangle, and heart. It's Regular only for now, installed with
npm i geist, and Thin to Black weights are in progress. (Vercel on X) - Figma agent + Google Drive – connect Drive so the Figma agent can read Docs, Sheets, and Slides, pull in context, and create new files from Figma. It's on Starter plans and up.
- Aura Autopilot (open source) – Ramen's playtesting agent for Unreal, Unity, and Godot reads your game's source, writes a control harness, and plays with real input. It works through Claude Code or Codex, under Apache 2.0. (Aura on X)
- &Tutor from humans& – a tutor model trained by teaching Persimmon, a simulated learner, and rewarded by how much the student learns. It's available via API. It leads TutorMoments under a plain prompt, though that benchmark uses an AI student and an LLM judge, not real learners. (humans& on X)
- TBC mischief 1 – an image-to-video model built on Wan 2.2 and optimized with measurements from living neurons. The neurons stay in the lab. It's about 5x faster than the base model and costs from half a cent per second on fal and OpenRouter. (Alex Ksendzovsky on X)
- PlanAssist – DeepMind's Gemini-based tool for UK planning officers, which it says cuts time on householder applications by up to 58%. It's expanding to 12 more councils.
What I'm Reading
- Cisco hands a 27-year-old unlimited AI tokens to refactor legacy code – the goal is 98% of Cisco's legacy code in 24 months with a two-person team, and he has four weeks to prove it works
- Harris on shipping to Instagram with agents – a grounded reply on why "make this kernel faster" beats "optimize my entire system," and why verifying user-facing features is still human work
- Agent frameworks vs. harnesses – a short, useful way to decide between importing a framework and starting from a working harness like Claude Code or Codex
That's it for today. 👋