A clearer view of AI.
Submit for reviewSubscribe

WEEKLY NEWSLETTER

The AI Insider

A weekly round-up of the AI tools worth your time, and the hype to ignore.

Unsubscribe at any time using the link in each newsletter.

THE INDEPENDENT AI.

AI NEWS FOR BUILDERS · 16 STORIES

AI news, 18 November 2026

From the community

Google lets you teach robot arms by chatting in Antigravity

Google DeepMind has released a robotics plugin for Antigravity that runs on Gemini 4 Hydrogen. You connect a supported robot arm, build and test a scene in MuJoCo simulation from a photo or a prompt, then run the task on real hardware using plain conversation. Google's demo goes from a whiteboard sketch to a working conveyor sorter in about an hour.

Why it mattersSmall manufacturers and labs could prototype simple automation such as pick-and-place or sorting without specialist robot programming. Google says Hydrogen matches its larger Argon model on these tasks at lower cost; that is its own claim.

Sourcesx.com post ↗

Google's Apigee turns existing company APIs into agent tools

Apigee Hybrid 1.17 now supports MCP, the standard way AI agents call tools. Companies can expose the APIs they already run, on their own servers or across clouds, as governed MCP tools without writing or maintaining custom MCP servers.

Why it mattersFirms that already manage their APIs in Apigee can let AI agents use internal systems while keeping the same access controls and monitoring, instead of building one-off connectors.

Sourcesx.com post ↗

ChatGPT for Sales links your CRM and sales tools

OpenAI has launched ChatGPT for Sales: a new Sales plugin plus more than 30 new and updated partner plugins, including Salesforce, HubSpot, ZoomInfo, Clay, Rox, Attio, Gong and Gamma. The idea is that ChatGPT can pull customer context from those tools and help with the admin around deals.

Why it mattersSales teams already using ChatGPT can bring their CRM, call notes and prospect data into one place instead of copying between apps. Check what data each plugin can see before connecting it.

Sourcesx.com post ↗

Tencent releases Hy4 flash as a free open-weight model

Tencent has published the weights of Hy4 flash, the low-cost everyday model it launched on 16 Nov, so anyone can download and run it. It is a large mixture-of-experts model (224B parameters, 12B active per answer) with a 1M-token context that handles text and images, and community GGUF versions for local tools are already out. Tencent says the full Hy4 model is coming next.

Why it mattersBusinesses that need to keep data in-house or cut API costs get another strong open model to host themselves. Tencent's earlier performance and cost comparisons for Hy4 flash are the company's own claims, so test it on your own work first.

Sourcesx.com post ↗

Claude Code's new Code Mode runs several tools in one step

Anthropic's developer account announced Code Mode, an experimental option in Claude Code. Instead of asking for one tool call at a time and reading each result, Claude writes a short JavaScript program that runs several commands, feeds one result into the next and returns only what it needs. You switch it on by starting Claude Code with CLAUDE_CODE_EXPERIMENTAL_CODE_MODE=1 (version 2.1.323 or later). Anthropic's docs say it only works in an interactive terminal session with a Claude subscription or Console account, not through Bedrock, other clouds, gateways, the VS Code panel or the Desktop app, and that it may change or be removed. Anthropic's developer account also announced a 'Latest' release channel you can opt into in the Claude Desktop app, and a Linked billing beta that lets users of tools like Factory, Linear and Ramp pay for Claude API usage through their own Claude Platform organisation without sharing an API key.

Why it mattersFewer back-and-forth steps can make long agent jobs faster and cheaper, which is where most coding-agent bills go. It's experimental, so try it on a side project first; if you manage a team, the docs explain how to block it in managed settings until you're ready.

Sourcesx.com post ↗ · code.claude.com ↗

Microsoft's Copilot Autopilot keeps working in the background, plus seven new MAI models

At Ignite on 17 November, Microsoft introduced Copilot Autopilot, an always-on agent that works across your email, meetings, chats and files to follow up, run recurring tasks and keep projects moving without waiting for a prompt. Microsoft says it has its own identity, memory and tools, with scoped, auditable access, and it's rolling out first to customers in Microsoft's Frontier early-access programme, who also need Copilot, Agent 365 and usage-based billing. Microsoft AI also launched seven of its own MAI models for coding, security, images, live voice, transcription and decision routing, which it says deliver frontier performance at a fraction of today's prices; those benchmarks are Microsoft's own, and several models are still in preview. A new MAI Licensed Weights option lets trusted customers customise a frontier MAI model and run it in their own Azure environment, in line with Microsoft AI's pitch that companies want control of their own model weights.

Why it mattersIf your business runs on Microsoft 365, an agent that acts without being asked raises new questions: what it can read, what it can send and who checks its work. Ask your IT partner how Autopilot's permissions and audit logs work before you join the early programme, and treat the MAI price claims as something to test against what you already pay for.

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

Gemini's Tasks agent is now free in six countries, but not the UK

Google has made Tasks in Gemini, its agent for longer projects and recurring jobs that used to be called Spark, free for eligible users in the US, Japan, South Korea, Canada, Australia and India, running on its new Gemini 4 Hydrogen model. You flip the home-screen toggle from Chat to Task and say what you need; free users can schedule a task up to once a day. US users also get extras such as having Gemini call businesses, pay with Google Pay, and take tasks by email or text. Google's help page says Tasks aren't available in the UK, the European Economic Area or Switzerland for now, even with a paid plan, and only work with personal accounts, not work or school ones.

Why it mattersFor UK and EU readers there's nothing to switch on yet. If you have staff or customers in the six countries, it's a free way to test hand-off agents on personal admin, but keep it away from work accounts, which Google doesn't support.

Sourcesx.com post ↗ · blog.google ↗

Google AI Studio now builds Android apps you can install on your own phone

Google AI Studio can now turn a single prompt into a full Android app with a database and Google sign-in through Firebase, call Gemini without exposing keys inside the app, run it in a faster in-browser emulator with sound, and publish it to your personal Google Play library so you can install it on your phone without a developer account. Google says personal library publishing is rolling out first in the US. Its official account also announced custom domains for web apps built in AI Studio Build. I watched Google's demo video; see Worth testing below.

Why it mattersIt's the quickest route yet from an idea to a private app on your own phone, which is ideal for testing an internal tool or a customer prototype before you pay a developer. Personal library apps are for you, not the public Play Store.

Sourcesx.com post ↗ · developers.googleblog.com ↗ · x.com post ↗Fact Hype or Fact? checkSteps to try, cost and caveat ↗

Perplexity launches one API and one bill for every major AI model

Perplexity launched its Orchestrator API, which gives developers models from OpenAI, Anthropic, Google and xAI plus open-weight models such as GLM, Kimi and DeepSeek through one API key and one bill. It accepts the OpenAI and Anthropic request formats, so switching is mostly a change of key and address, it retries a failed request on another copy of the same model, and Perplexity says there's no platform fee on top of credits. Plain-language routing rules are in preview. Perplexity Computer also gained a Multiagent effort level for Max subscribers on the web, where a lead agent splits work across up to 16 agents on different frontier models that check each other's results.

Why it mattersIf your product calls several AI providers, one bill and automatic failover can cut admin and outages. Compare it with OpenRouter, which Hugging Face switched to this week, on price and on which models each one actually offers.

Sourcesx.com post ↗ · perplexity.ai ↗ · x.com post ↗

Mistral gave European developers three free months of Vibe Pro, and the codes ran out

Mistral offered developers in Europe three months of Vibe Pro free with the code LECHONK, a plan that normally costs EUR 17.99 a month in France, first come, first served; it later said the codes had run out for the day. Mistral claims its Vibe coding agent with 'Le Chonk', its nickname for its open-weight Large 4 model, is the world's best open model and agent pairing, and better than GLM and DeepSeek at coding. On Artificial Analysis's Coding Agent Index, as reported by Sovereign Magazine, Vibe with Large 4 scores 41 against 40 for GLM-5.3, but Moonshot's open-weight Kimi K3.1 Pro scores 49, and Vibe is the cheapest per task of those listed. Mistral doesn't say whether the UK counts as Europe for the offer. Our verdict: partly fact.

Why it mattersIf you want a European coding agent for data or sovereignty reasons, Vibe is now competitive and cheap to run, but 'world's best open' overstates it. Watch for the offer to return if you missed it, and test it on your own code before committing.

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

OpenAI's API dashboard now shows spend, rate limits and errors by key

OpenAI's developer account says the API usage dashboard now makes it easier to track usage and spend, check rate limits and rate-limited requests, and explore errors broken down by model, project or API key.

Why it mattersIf a developer or agency runs AI features for you on OpenAI, ask them to check this dashboard monthly: it's the simplest way to spot a runaway cost or a failing feature before your customers do.

Sourcesx.com post ↗

Isomorphic Labs raises $5 billion for AI-designed medicines

Isomorphic Labs, the Alphabet drug-discovery company led by Demis Hassabis, raised $5 billion in a round co-led by Greenoaks and Sequoia. Some posts called it a Series A, but the company and TechCrunch describe it as a Series C. The company is preparing to take its drug candidates into clinical trials. Shared by @kimmonismus.

Why it mattersNot a tool you can use today, but a sign of where the biggest AI money is heading next: science and health, not just chatbots. If you work in life sciences, expect more AI-first competitors and partners.

Sourcesx.com post ↗ · techcrunch.com ↗

Agent Substrate: an open-source way to run huge numbers of agents cheaply

Agent Substrate is a free, open-source system for running very large numbers of AI agents on Kubernetes. Because agents spend most of their time waiting, it puts idle ones to sleep and wakes them on request onto a small pool of servers; the project claims they resume in under half a second with about ten times more agents on the same hardware, which are its own figures. Each agent runs in its own sandbox, and API keys are kept outside it, so a hijacked agent has nothing to steal. The GitHub repo, checked tonight, is Apache-2.0 licensed with about 7,300 stars. Shared by @Saboo_Shubham_.

Why it mattersIf you plan to give every customer or employee their own agent, the server bill and security are the hard parts. This is one for your developers to evaluate; it needs Kubernetes skills.

Sourcesx.com post ↗ · github.com ↗

Chrome's Modern Web Guidance teaches coding agents web best practices

Google Chrome's Modern Web Guidance is a free, open-source set of agent skills that gives coding agents expert-checked advice on the things they often get wrong, such as keyboard focus, reduced motion, layout shift, dark mode and safe HTML. Version 1.0 came out on 10 November with Firefox and Edge now involved, and installs with npx modern-web-guidance@latest install or from the plugin stores of tools like Claude Code, Codex and Cursor. On 17 November @ctatedev showed a new agent-browser skill that looks up the guidance, applies it to your code and then checks the result in a real browser (npx skills add vercel-labs/agent-browser).

Why it mattersIf you build your website or app with an AI agent, this is a free way to make the result more accessible and faster without hiring a specialist to review every page.

Sourcesx.com post ↗ · developer.chrome.com ↗

Matt Pocock: give agents fewer choices, and a prompt to grade any new library

@mattpocockuk is experimenting with '/constrain': if too much choice makes agents worse, build a deliberately narrow version of the tools you like, enforced by strict lint rules, so the agent can only take the right path. His prompt of the day applies the same idea to choosing technology: ask the agent to research a library, find the hardest problem in your app it could help with, build a proof of concept in a draft pull request, verify it, then grade it on how many choices its API leaves open, type safety, testability, added complexity, docs and deployment changes, and give a recommendation. It's his method, not a measured result, and it follows Lauren Tan's 'nuance leads to slop' talk from yesterday's edition.

Why it mattersWhenever you're tempted by a new tool, a structured agent trial like this gives you evidence instead of hype, and the 'fewer choices' rule is a cheap way to get more reliable work from agents.

Sourcesx.com post ↗ · 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.