A clearer view of AI.
THE INDEPENDENT AI.

AI NEWS FOR BUILDERS · 15 STORIES

AI news, 25 October 2026

Kimi K3.1 Pro: a leaked brief alleges a paid, unlabelled influencer campaign; OpenAI is using discounts to win Anthropic's coding customers; How Anthropic keeps Claude Duet away from your passwords and money.

Top stories

Kimi K3.1 Pro: a leaked brief alleges a paid, unlabelled influencer campaign

A follow-up to yesterday's top story: screenshots of what's said to be a creator brief for Moonshot's Kimi K3.1 Pro show posts needing pre-approval, no 'paid partnership' label, no need to tag or link Kimi, and scripted lines such as '5x cheaper open-source Opus 5.5' and 'China is one month behind the US frontier'. The screenshots haven't been independently authenticated and Moonshot hadn't responded when we checked, so treat this as an allegation, not a proven fact.

Why it mattersSome of the 'this new model beats Claude' posts in your feed may be paid ads. Test models on your own work before switching, and if you ever pay creators to post, UK advertising rules (the CAP Code), the FTC in the US and X's own rules all require a clear #ad label.

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

OpenAI is using discounts to win Anthropic's coding customers

The Information reports that OpenAI is offering steep discounts to customers of Anthropic: Replit reportedly got more than 30% off a contract worth over $100M a year, and TechCrunch has confirmed Replit made OpenAI its default model, with Claude still available. At Rippling, OpenAI's share of coding work reportedly rose from 16% in May to 52% while Anthropic's fell from 57% to 18%; those figures come from Rippling's CEO via the report, not public data. @kimmonismus shared it.

Why it mattersAI model prices are now negotiable and the big labs are fighting for share. Don't lock yourself into one provider: ask for a discount at renewal, and build your workflows so you can switch models without starting over.

Sourcesx.com post ↗

How Anthropic keeps Claude Duet away from your passwords and money

Anthropic explained the guardrails on its Duet agent: approval cards plus a separate intent check that runs outside the model, actions it refuses even with your approval (trading, moving money between accounts, using someone else's account), and a classifier that sorts sensitive data into 'never share' and 'needs consent'. Passwords are only unlocked on an isolated cloud browser and sign-in tokens are added by a proxy, so the model never sees them, and purchases use one-time Link cards. These are Anthropic's own claims; only the penetration test was external, and its report isn't public.

Why it mattersThis is a ready-made checklist for any agent you connect to email, banking or client accounts. Ask the vendor two questions: does the model ever see our passwords, and what can't it do even if we approve it?

Sourcesx.com post ↗ · claude.com ↗

Big lab updates

OpenAI's Codex CLI 0.164 adds reusable prompt presets

The new /presets command saves up to 10 prompts you can reuse with one command, and the release also adds image previews in the terminal, sign-in for company MCP servers, credential masking and fixes for resuming long sessions. Codex lead Thibault Sottiaux (@thsottiaux) says the command-line tool has doubled in use since DevDay, which is OpenAI's own figure.

Why it mattersSave 'write my weekly client update' or 'review this proposal' once and reuse it in one step. It's the start of a personal prompt library inside your coding agent.

Sourcesx.com post ↗ · x.com post ↗

Rumours from the leak trackers: Gemini 4 'Carbon', a new Google image model and Qwen 4

@Mr_Salio spotted a Gemini 4 'Carbon' model with 3D output being tested on Arena, Google testing a new image model under a Nano Banana 2.1 label, and a Qwen 4 checkpoint. @testingcatalog's daily brief adds Cohere's Model Vault in public beta, 4K output from Seedance 2.5 for enterprise customers, and a report that Thariq Shihipar has left Anthropic. None of this is officially announced, so treat it as reports.

Why it mattersNothing to act on yet, but a Gemini that outputs 3D would matter if you make product or event visuals. Wait for the official launch before you test it.

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

Builder workflows

Mariah Brunner's AI motion graphics starter kit

Fact or hype? Mostly fact. We couldn't play the reel itself, so we checked the sites it names, and they're real and mostly free. Prompt Motion (promptmotion.org) is a free gallery of Claude Opus motion videos with the exact prompt behind each; React Bits (reactbits.dev) offers free animated React components you can use commercially in your own site or app but can't resell; and Eyecandy (eyecannndy.com) is a free library of visual techniques that gives you the right words for video prompts.

Savee is a design-reference board whose connector for Claude and Cursor (the Savee MCP) is only on paid plans, from $9 a month billed yearly. We left out MotionSites, which advertises on Prompt Motion, because most of its prompts are paid.

Why it mattersYou can make launch videos and animated pages without hiring a designer. Start free with a Prompt Motion prompt and React Bits components, and only pay for Savee if the connector saves you real time.

Sourcesinstagram.com ↗

New tools

HarnessRouter: one gateway to run Claude Code, Codex and other agents

An open-source gateway (Apache-2.0, about 3.4K GitHub stars, self-hostable with Docker) that puts 26 agent 'harnesses' behind one OpenAI-style API, so one agent can hand a task to another. The headline '99.8% lower cost, 3.2x faster' is the company's own benchmark, comparing its best and worst of 8 setups on a single task, and cloud plans run from $0 to $1,000 a month, per the company. The post shows no sponsorship disclosure and reads like a promo; see the demo under Worth testing.

Why it mattersMixing agents (one to analyse, one to build) without being tied to a single vendor is useful, but this is developer infrastructure, not a no-code tool.

Sourcesx.com post ↗Partly fact Hype or Fact? checkSteps to try, cost and caveat ↗

CLI-Anything: let coding agents drive your desktop apps

An open-source project from the University of Hong Kong's Data Intelligence Lab (about 56.8K GitHub stars, Apache-2.0) that wraps desktop apps such as Audacity, Blender and Obsidian in command-line tools, so agents like Claude Code or Cursor can operate them. The post resurfaces an existing project rather than announcing a launch.

Why it mattersYou could automate repetitive desktop jobs, like cleaning up podcast audio or tidying notes, without waiting for an official integration. You'll need some comfort with the terminal.

Sourcesx.com post ↗

Sutyo AI: a 'personal memory' app

A new iOS and Android app from Sutyo, based in Austin, that stores your voice notes, screenshots, links and documents, answers questions with links back to the original, and reminds you about follow-ups. It's very new, pricing isn't confirmed, and 'audio transcribed on the device and then deleted' is the company's own claim.

Why it mattersA place to dump voice notes after a call and later ask 'what did I promise this client?'. Get people's consent before you record them (UK GDPR).

Sourcesx.com post ↗

Skills & prompting

Report: Perplexity's 'Taste' skill is free until 31 October

@testingcatalog reports that Perplexity's 'Taste' skill is now in the skills marketplace for Perplexity Computer and free to use until 31 October. That comes from a leak-tracking account and we haven't tested the skill ourselves.

Why it mattersIf you already use Perplexity Computer, this is a free week to see whether a ready-made skill improves your results. Try it on one real task before the offer ends.

Sourcesx.com post ↗

Harness engineering paper: the 7 building blocks every coding agent shares

A Wavestone AI Lab paper studied the source code of 11 coding agents and found the same seven building blocks in each: the loop, the model layer, tools, memory, safety, orchestration and extensions. @Arindam_1729 summarised it.

Why it mattersA simple checklist for judging any agent product: before you trust it with client work, ask where its memory, safety rules and tools come from.

Sourcesx.com post ↗

Worth testing

AI search tip

Ask each AI assistant who it recommends in your area: they rarely agree

Yext Research looked at 491,972 local searches from February to August 2026 and found that ChatGPT, Gemini, Claude and Perplexity named the same top business only 4% of the time, and picked four different businesses 48% of the time. When they did agree, the winner's own website was cited 91% of the time. Each engine leans on different sources: Gemini mostly on business websites (41.9%), OpenAI on directories (40.9%) and Anthropic on review sites (20.5%). This shows correlation, not cause, and Yext sells listings management, so weigh it accordingly. Try it: keep clear service, area, price and FAQ pages on your site and consistent listings on the main directories and review sites, then once a month ask all four assistants for the 'top 10 [your category] near [your area]' and note where you appear.

That's it for today.

Sourcesyext.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.