AI on X Summary (October 31, 2026)

Here are some interesting things that happened on my X timeline in the last day: Anthropic reshuffled its subscription names and lifted a usage cap, Google says it ran a prompt on a TPU in orbit, and a wave of early evaluations for Gemini 4 Carbon arrived.

What got attention

Anthropic renames Max and lifts the 5-hour limit. According to @claudeai, the $100 plan formerly called Max 5x is now just Max, and the $200 Max 20x plan becomes Ultra, with no action needed from existing subscribers. The bigger news for heavy users is that the 5-hour limit is being removed for verified Ultra accounts, rolling out starting that day. @bcherny and @pablostanley both singled out the limit change as the headline. This is a company announcement, so how it plays out in practice is still to be seen.

A prompt answered from space. @sundarpichai reports that Project Suncatcher, a prototype satellite launched a few weeks ago, received a command, booted its TPU, ran a prompt and sent back a complete answer: a haiku. @alanchen adds detail, saying the satellite carries four TPUs but can power only one at a time, for about 20 minutes, on roughly a kilowatt in total. @_philschmid asked, as a question rather than a claim, whether this makes Gemma 4 the first LLM to run in space.

Gemini 4 Carbon evaluations. @AravSrinivas says Perplexity’s early evals show it as the strongest model they have tested for long-form research, with a new top score on their WANDR benchmark. @mercor reports it debuting first on APEX-Agents, ahead of Gemini 4 Argon and Claude Sonnet 5.5. @sumitbagri, who worked on the model, highlights Terminal-Bench results and a very low rate of cheating attempts. Note that these come from the model’s team and from partners preparing to ship it, not from independent testers.

Codex quality-of-life updates. @thsottiaux announced account switching on mobile, with desktop to follow, and hinted at usage resets over the next two days. @Dimillian celebrated Live Activities for Codex in the ChatGPT iOS app, while @athyuttamre acknowledged a dictation regression in the latest desktop app that the team is working on.

Automated red-teaming. @AnthropicAI shared research in which thousands of Claude agents were set to work on safety-testing tasks, as a way to get past the limit of how much time human researchers have for red-teaming.

Practices worth copying

• Use a second model as a design reviewer. @thdxr says their team calls a Gemini subagent for hard architecture problems, system design questions and simplification work. Mixing models for different kinds of thinking is cheap to try.
• Send the grunt work to cheaper models. @mylifcc describes a Spotify plugin called shunt that routes file reading and boilerplate test writing away from the most expensive model in Claude Code, reportedly cutting token use by about 90%. The figure is as reported in the post.
• Give the agent full context, then let it run. @geniusvczh says they handed an agent all the background on a bug nobody knew how to approach; after 36 hours of exploration it opened a pull request with a three-line fix.
• Fix CI before it becomes the bottleneck. @owengretzinger argues that faster AI-assisted coding moves the bottleneck to CI, and says they cut theirs from 40 minutes to 2. The details are in their thread.
• Make proactive agents quiet by default. @gabemulley says early versions of Claude Duet’s proactivity notified too often about things they already knew, so the team now favors fewer, higher-value nudges.
• Know whether you need a framework or a harness. @Saboo_Shubham_ separates libraries you import to assemble an agent from harnesses that already decide the loop, context and checks.

Doubts and open questions

• A post quoted by @geniusvczh warns that in code review, developers increasingly cannot explain the AI-written code they submit, for example why a given transaction isolation level was chosen.
• @Zen_with_AI lists OpenAI products that lasted under a year, and says Prism, its writing space for scientists, will shut down on January 1, 2027, with data deleted afterward. A reminder to keep exports of anything you build on a new tool.
• @Zen_with_AI points to a chart showing AI Overviews appearing on far more Google searches this year, alongside lower click-through for the top organic result.
• @WSJ reports that AI is starting to speed up AI research itself, and that this is unsettling people.
• On a lighter note, @reach_vb says solving a few coding problems by hand now feels surreal.

Takeaways (my own synthesis)

• Usage limits are becoming a product feature in their own right; both Anthropic and OpenAI spent the day talking about caps and resets.
• Multi-model setups are normal now: one model for code, another for design review, a cheaper one for routine I/O.
• Long, autonomous agent runs can pay off, but only if you can still explain and review what comes back.
• Treat launch-day benchmark numbers as claims until outside testers weigh in.