A clearer view of AI.
THE INDEPENDENT AI.

AI NEWS FOR BUILDERS · 11 STORIES

AI news, 26 October 2026

Kimi says sorry for K3.1 Pro rate limits, but not for the influencer brief; CapCut is coming to ChatGPT (and part of it already works, if you're on the right setup); Nanbeige4.3-33B: a free open model aimed at office and coding agents.

Top stories

Kimi says sorry for K3.1 Pro rate limits, but not for the influencer brief

Moonshot AI apologised for the rate limits and wasted quota on its Kimi K3.1 Pro subscription plans and is giving every subscriber a one-time reset that restores 25% of their monthly quota (about a week of use). It also says it has cut unnecessary sub-agent calls and shortened outputs so quota lasts longer. @kimmonismus pointed out that the statement says nothing about the leaked creator brief we covered yesterday, which allegedly asked influencers for undisclosed paid posts.

Why it mattersTwo lessons from one week: a cheap model can still cost you time if its plan throttles you mid-task, so test usage limits before you move work onto it; and treat glowing posts about it with caution until the disclosure question is answered.

Sourcesx.com post ↗ · x.com post ↗

CapCut is coming to ChatGPT (and part of it already works, if you're on the right setup)

At its CapCut World LA event on 24 October, CapCut (ByteDance's video editor) announced a plugin that's 'coming soon' to ChatGPT: start with an idea in the chat, generate and edit a draft video with CapCut there, then open it in CapCut's editor to finish and publish. There's no launch date, region list or price for that plugin yet. What exists today is an official 'CapCut × Codex' workflow for the ChatGPT desktop app on PC only (not web or mobile); CapCut calls it free to start but says availability varies by app version, account and region, and third-party reports say it's outside the US for now. CapCut's AI generation still runs on credits (Pro is USD 19.99 a month with 1,200 AI credits, US price). Many of the demo posts on X look like coordinated promotion.

Why it mattersIf you make short social videos, this could cut out the copy-paste between your script and your editor. For now treat it as an announcement: if you're on the ChatGPT desktop app on a PC outside the US, try the Codex workflow on one simple product video, and read every step it asks to run before you approve it.

Sourcesx.com post ↗ · capcut.com ↗Partly fact Hype or Fact? check

Nanbeige4.3-33B: a free open model aimed at office and coding agents

Chinese lab Nanbeige released Nanbeige4.3-33B on Hugging Face under the Apache 2.0 licence, so businesses can use and modify it freely. It's built for agent work, from coding to 'cowork' tasks like automation and planning, and handles very long documents (about 512,000 tokens). The lab says it beats similar-sized Qwen and Gemma models on most agent benchmarks, but those are all its own tests; Qwen still edges it on one major coding benchmark. @kimmonismus flagged it.

Why it mattersIf you need an AI agent that runs on your own servers so client data never leaves, this is a new candidate to compare. Be realistic: the full model is about 66 GB and the lab's setup uses four GPUs, so it's a job for a hosted server, not a laptop.

Sourcesx.com post ↗ · x.com post ↗ · huggingface.co ↗

Big lab updates

Rumour roundup: what might land this week (none of it confirmed)

• Anthropic: @ChrisGPT says sources tell him a model called Fable 5.5 has been rolled out to enterprise clients including JP Morgan and Meta, and @kimmonismus expects a public release within days, linking it to a reported Anthropic IPO around 9 November. Neither the rollout nor the IPO date is confirmed by Anthropic: x.com and x.com

• OpenAI: two mystery models ('shooting-star-red-alpha' and 'shooting-star-green-alpha') appeared on the LMArena testing site on 23 October, and @Mr_Salio says they're OpenAI's. Separately, a model called 'GPT-6.1 Terra' is said to be showing up in partner and Microsoft test configs: x.com and x.com

• Codex: OpenAI's 28 days of Codex improvements end this week, and Codex lead Tibo (@thsottiaux) promised a 'fun week ahead', which has fuelled talk of a new model: x.com

Why it mattersIf you're about to commit to a plan or a provider, a week's wait could get you a newer model or a better deal. Don't rebuild anything on a rumour; wait for the official release notes and pricing.

Sourcesx.com post ↗ · x.com post ↗ · x.com post ↗ · x.com post ↗ · x.com post ↗

Agents & use cases

Don't wait for your agents; let them wait for you

Developer Chris Tate (@ctatedev) backed a widely shared piece by Thorsten Ball arguing that you should stop worrying about how fast an agent replies and instead hand it bigger tasks that run in the background while you do something else.

Why it mattersThis is the mindset shift that makes agents pay off: batch up a meaty job (research a market, draft five proposals, clean a spreadsheet), set it running, and check back later, instead of chatting with it line by line.

Sourcesx.com post ↗ · x.com post ↗

People are using Claude Code to strip ads and bloatware off their smart TVs

Several posts show Claude Code connecting to Android or Google TVs over the home network and switching off ad and tracking apps, with @kimmonismus calling it proof of why local AI matters. One step-by-step guide says it works on Android TV and Google TV sets, but not Samsung or LG, which run their own systems. We haven't tested it, and changing system settings on a device carries some risk.

Why it mattersBeyond the TV, it's a good example of an agent doing a fiddly technical job for a non-technical person. Ask yourself which annoying admin task in your business an agent could talk you through the same way.

Sourcesx.com post ↗ · x.com post ↗

Builder workflows

Matt Pocock: a 'software factory' running on GitHub Actions

TypeScript teacher Matt Pocock (@mattpocockuk) says his new course ends with a basic 'software factory': automated agents that pick up tasks, write code and get it reviewed, all running on GitHub's built-in automation, plus a new skill to help you plan your own. In a separate post he suggests that some code changes could be approved for merging automatically if they meet very strict rules, such as only reorganising code and never reducing test coverage.

Why it mattersEven if you don't code, the idea is useful: decide which kinds of work are safe for automation to finish end to end, and which always need a human check. It's the same rule for agents writing your emails or your code. It's a paid course, so this is Matt promoting his own product.

Sourcesx.com post ↗ · x.com post ↗

Worth testing

AI search tip

Answer the follow-up questions, because Google's AI searches for them too

Google's own guidance says AI Overviews and AI Mode may use 'query fan-out': they run several related searches across subtopics before writing one answer, so a page can be cited for a side question even if it doesn't rank for the main one. For a small business, that means each key service page should also answer the obvious follow-ups (price range, who it's for, how long it takes, how it compares with the alternative), with clear headings and internal links between those pages. Google also says a page only qualifies if it's indexed and allowed to show a snippet, so check you haven't added nosnippet or a tight max-snippet tag by accident. To check it's working, use the URL Inspection tool in Search Console on your key pages, watch the Performance report (AI features are counted in the 'Web' search type), and try a few follow-up questions in AI Mode to see if you're linked.

That's it for today.

Sourcesdevelopers.google.com ↗ · x.com post ↗

How we knowCompany figures are labelled as their own claims; anything we couldn't verify is labelled as a report or a claim.