AI on X Summary (October 29, 2026)

Published October 29, 2026 · 4 min read

Here are some interesting things that happened on my X timeline in the last day: OpenAI doubled its Codex usage limits, Anthropic raised the Claude API request size cap, and Google’s new Gemini 4 models drew a lot of reactions.

What got attention

Usage limits came back. OpenAI announced it is doubling usage limits for its Pro tiers and Business Premium, effective immediately. @ajambrosino says the feedback worked and that a sprint to improve the app’s simplicity and performance has started. @angelbrodin repeats that limits are not being reduced. Reaction was mixed: @Kay2289123 argues the limits were halved a month ago and now simply restored, so nothing improved. This is a company announcement, not independent confirmation.

Claude API got more room per request. According to @ClaudeDevs, the maximum request size for Messages and Token Counting went from 32 MB to 128 MB, which matters for long computer-use sessions and large PDFs. A follow-up post says the per-image limit rises from 2000 px to 3000 px for requests with more than 20 images, and that the 128 MB limit applies only on the first-party API (@ClaudeDevs). @Zen_with_AI explains why it matters: a high-resolution screenshot is a megabyte or two in the request but only a few thousand tokens to the model.

Gemini 4 chatter. @BHolmesDev calls Gemini 4 Argon the best copywriter they have seen, citing tone matching and default voice. @GoogleAIStudio announced XR app building from a single prompt in AI Studio, and @joshwoodward says Notebook unlocked over 700 free textbooks from publishers. Both are announcements from Google.

Open models and security. @Zen_with_AI reports that Mistral Large 4, a 1T-parameter mixture-of-experts model with about 52B active parameters, can run locally with Unsloth’s dynamic quantization, citing roughly 233 GB for the 1-bit version. The same account relays a claim that an unreleased Anthropic model chained a Linux kernel TCP vulnerability into remote root execution in about 53 minutes, and urges Ubuntu users to patch (post). That is a secondhand claim on one post. @Yuchenj_UW jokes that security researchers will soon out-earn AI researchers.

New Claude products. @pablostanley previewed Claude Duet with a waitlist, and @ahuja_karan says Tasks and Channels are central to it. Both rest on a design preview, not a release.

Practices worth copying

  • Constrain what the AI is allowed to write. @vikingmute suggests telling an AI to avoid useState in React, since it tends to pile on local state that drifts out of sync. They also recommend a team’s essay on banning useEffect.
  • Use agents as a parallel review team. @mylifcc describes an open-source Cloudflare security-audit skill that runs several coding agents in parallel to find bugs, verify attack paths and write a report.
  • Combine tools for AI-made video. @theEarlyPanda says striking Claude-made videos mostly come from orchestrating Blender, Three.js and p5.js.
  • Try AI for creative editing with public-domain footage. @lukeigel says Kino, which drives Claude Code and Codex to edit video, works with archival material and tracks sources.

Doubts and open questions

@juokaz reports DoorDash saying personal agents like Muse have produced only “several hundred orders” in total. @athyuttamre says an agent hallucinated a legal case and a Zoom hearing against them, a reminder that agents are early. On the money side, @oran_ge worries AI investment is reaching its limit and the return on it is hard to justify. @xiaojingcanxue notes that nearly half of Anthropic’s revenue, per its prospectus, comes from big IT firms that are also laying off staff. Each of these rests on one post.

Takeaways (my own synthesis)

  • Usage limits are now a competitive feature; expect more back-and-forth.
  • Bigger request limits make screenshot- and PDF-heavy agent workflows more practical.
  • Constraining the agent (banned patterns, parallel reviewers) looks more reliable than trusting it with open-ended work.
  • Consumer agents still have trust problems; the evidence this week is thin and anecdotal.
← Back to home