AI on X Summary (October 28, 2026)

Published October 28, 2026 · 4 min read

Here are some interesting things that happened on my X timeline today: a new Google model, OpenAI folding API credits into ChatGPT plans, and a lot of talk about MCP.

What got attention

Google’s Gemini 4 Carbon. The launch dominated the AI list. @OfficialLoganK announced it two weeks after the previous Gemini 4 release and claims leading results on cybersecurity and long-horizon tasks, first place on the Artificial Analysis index and better coding. @demishassabis repeated the benchmark claim. The outside number came from the leaderboard account @arena, which says the model debuted first in Text Arena, second in Agent Arena and fourth in the WebDev coding arena. One post relays Chrome team testing in which it beat another lab’s model by more than ten times on a security task (@DynamicWebPaige), but that is second-hand. A Chinese-language reaction (@qingzhoucloud) reads the launch as Google regaining ground and ranks OpenAI last of the three big labs.

Google also put a usage figure on the table. Relaying its earnings post, @Saboo_Shubham_ says Google’s own model APIs now process close to 32 billion tokens a minute.

OpenAI puts API credits inside ChatGPT plans. @OpenAIDevs says eligible Pro and Business subscribers can claim up to $500 a month in API credits for building apps and agents. A follow-up gives the split: Pro tiers get $100, $200 or $500, and Business gets $20 per Standard seat or $100 per Premium seat, pooled up to $500 per workspace. @romainhuet teased it on stage at a Paris developer event after an audience question. Usage limits were still on people’s minds: @FiniYang asked whether the Codex Pro 200 allowance really would not be halved, and @thsottiaux said a rumored merge of Chat and Work in ChatGPT is not rolling out yet.

Agents are blurring who counts as technical. One of the most-liked posts of the day, from @trq212 with about 19,000 likes, observes that engineers fear their skills are now worthless while non-technical people wish they knew more so they could direct agents better. A Chinese-language reply (@uudonX) says that matches what they see: non-technical people learning to build, including some who had given up on programming.

MCP is becoming a distribution channel. @MishaalRahman says Google is launching Android MCP on its AppFunctions API, so an app can let Gemini on Android pull information from it or act inside it. @daluoseo reports a similar Connected Apps program that lets developers put a product inside Gemini by writing an MCP server, said to reach Search in a few weeks and open only to US developers for now. Gemini also gained a Strava connector that turns workout history into a coaching plan (@joshwoodward), and Gemini Nano 4 lets Android apps add an offline agent through ML Kit (@MishaalRahman).

Practices worth copying

  • Give your pull requests a minder. @btraut recommends a Codex skill that handles review comments, merge conflicts and CI failures by itself and only interrupts you when it needs you. The post points to the openai/codex repository for an example.
  • Make your product something an agent can call. @daluoseo advises SaaS builders to wrap their core feature as an MCP server, submit it to the MCP directories, and write documentation in plain language, because an AI that cannot understand your docs will not recommend you.
  • Share session IDs when usage looks wrong. @bcherny says that after users shared theirs, the team found several smaller problems that made some sessions use tokens faster than they should, and is rolling out fixes.
  • Reach your agent where you already are. @rileybrown describes pairing OpenAI’s “dot” assistant with an always-on Mac mini and controlling the computer entirely by voice. A post from @thsottiaux says you can now text your dot, hand it tasks or things to watch, forward it items and leave voice notes.
  • Learn the loop, not just the model. @0xkichiro points to a free two-hour lecture by Andrej Karpathy that moves from agents through harnesses and loops to graphs. I have not watched it, and the post is promotional, so judge it yourself.

Doubts and open questions

  • Output nobody reads. @coolXiao notes that an AI tool can produce a 1,500-word competitor brief, tagged and linked, in 45 seconds. The catch is that you never read any of it.
  • AI in the wrong place. @zzxwill complains that Google Maps now shows an AI summary and reviews above the location itself.
  • AGI skepticism. @0xNing0x argues that current methods only compress human knowledge and will not produce a general intelligence, and that AGI promises have repeatedly been missed or redefined.
  • Vendor-reported results. The benchmark wins, the tenfold security figure, and @aastroza reporting that ProChile cut a two-day task to ten minutes with Codex all come from the companies or their own events.

Takeaways (my synthesis, not from the posts)

  • Distribution is moving to agent surfaces. Gemini, Android and ChatGPT each want your product exposed as a tool, and an MCP server is the common way in.
  • Trying agents keeps getting cheaper. Bundled API credits remove one more reason not to build something.
  • The tasks people hand to agents first are the dull ones around the work: pull request upkeep, CI, watching for changes.
  • Keep a human in the reading loop. The fastest way to produce text is also the fastest way to end up with documents nobody has read.
← Back to home