The new features, tips and tricks for ChatGPT, Claude, Gemini and Meta Muse that are worth your time this week, plus the industry news behind them and prompt packs you can copy. Each item explains why it matters, who it is for and which plan you need, in our own words, with a link to the original source.
A prompt creator shared a structured prompt for Nano Banana 2.1 that is written like a SQL query. It asks for a 2x2 grid of four cities, each shown as a tiny spherical planet resting on an open palm, with the city's main monument on top, dense blocks around it, highways wrapping the sphere, local trees, and a tilt-shift, bokeh look.
Why it matters: Writing a prompt as a list of named fields makes it easy to swap one part (the city, the lens, the trees) and keep everything else the same.
Good for:
Designers and hobbyists who like repeatable image series.
Plan & availability:
Nano Banana 2.1 in the Gemini app, AI Studio or the Gemini API.
Note: reported by users on X; we haven't independently confirmed it works in the current version.
Developer and creator Theo posted that Gemini is about 50% more expensive than it should be for his workload because caching isn't working well, citing a 27% cache miss rate. The post drew a lot of attention from other developers.
Why it matters: If you rely on prompt caching to keep costs down, a high miss rate quietly raises your bill.
Good for:
Developers running high-volume Gemini API workloads.
Applies to:
Gemini API.
Note: reported by users on X; we haven't independently confirmed it works in the current version.
Boris Cherny, who leads Claude Code at Anthropic, says Claude Slides is now how he makes decks, including the device-frame mockups he used to search for by hand. Slides lives in the Claude app alongside Claude Docs and Claude Design, so you can draft a one-pager, turn it into a deck and mock up a matching visual in one conversation.
Why it matters: You can go from notes to a finished deck with mockups without leaving Claude or hunting for device frames.
Good for:
Founders, PMs and marketers who make decks often.
Plan & availability:
Claude app. Check Anthropic's plan page for which plans include Slides.
Boris Cherny says Claude Code Review runs on all of his pull requests and catches almost every bug before it ships, and that it is part of how his team moves fast without lowering the quality bar. It is worth switching on if you merge a lot of agent-written code and want a second reviewer that never gets tired.
Why it matters: Agent-written code ships faster than humans can review it. An automated reviewer on every PR catches bugs before they merge.
Good for:
Engineering teams that merge a lot of AI-written pull requests.
Plan & availability:
Claude Code; setup and plan requirements are in Anthropic's Claude Code docs.
OpenAI upgraded GPT-6 Astra Pro in ChatGPT on October 16 with faster responses and better use of plugins and tools. Multiple agents can now work on different parts of one task in parallel, and you can follow each agent's progress and open finished parts to review them. It is aimed at Pro users with big, splittable jobs like research across many sources.
Why it matters: Big research or comparison jobs finish faster when separate agents take separate parts, and you can check each part as it lands.
Good for:
Pro users doing multi-source research, vendor comparisons or long reports.
Plan & availability:
Rolling out to eligible ChatGPT Pro users; managed workspaces depend on admin settings.
OpenAI Developers shared how OpenAI's own engineers use Codex to dig through their codebase, find what needs attention and work through the maintenance backlog. It is a practical pattern for the boring work (dead code, outdated dependencies, flaky tests) that never gets prioritised.
Why it matters: Maintenance work rarely gets prioritised. Letting Codex find and size it turns it into small PRs you can merge between features.
Good for:
Engineers and tech leads with a growing backlog of cleanup tasks.
Plan & availability:
Codex in the ChatGPT desktop app, CLI or IDE extension on plans that include Codex.
An OpenAI engineer used his dot to triage a 3am on-call incident: the dot investigated the issue and started a Codex thread, and he steered the fix from his phone in the ChatGPT app. It shows how dots and Codex hand work to each other so you do not need a laptop to start an investigation.
Why it matters: Starting an incident investigation from your phone saves the minutes it takes to get to a laptop at 3am.
Good for:
On-call engineers and small teams without a dedicated SRE.
Plan & availability:
dots are rolling out to eligible Pro and Business Premium users (not in the EEA, Switzerland or UK at launch).
OpenAI Developers published a guide on giving GPT-Live-1 apps a voice that fits: you can shape a voice with a prompt or choose from new international preset voices. Useful if your voice agent sounds generic or wrong for your audience.
Why it matters: A voice that matches your brand and market makes a voice agent feel trustworthy instead of robotic.
Good for:
Developers building voice agents with GPT-Live.
Plan & availability:
OpenAI API (GPT-Live-1). See the OpenAI Developers guide for voice options.
With Gemini 4 Argon, Canvas can turn a dense PDF into an interactive, animated web app you can explore at your own pace. It is a great way to make a research paper, manual or report actually readable.
Why it matters: An interactive explainer is much easier to learn from than a 40-page PDF, and you can share it with a team.
Good for:
Students, researchers and anyone onboarding onto dense documents.
Plan & availability:
Gemini app with Canvas; Gemini 4 Argon access depends on your Google AI plan.
As of October 16, Google AI Pro subscribers can build with Gemini 4 Argon in AI Studio Build and the Playground. Before that it was limited to paid API and Ultra subscribers. It is the cheapest way to prototype small apps with Argon if you already pay for AI Pro.
Why it matters: AI Pro subscribers can now prototype with Google's top model without paying for API usage separately.
Good for:
Indie developers and tinkerers who already pay for Google AI Pro.
Plan & availability:
Google AI Pro, AI Studio Build and Playground (since Oct 16, 2026).
Google DeepMind ran four Gemini 4 Argon agents in a sandbox, using only code and text. They wrote 2,783 programs, split up jobs over chat, shared notes and debugged together to beat Minecraft in 27 hours with no human help. A good read on how multi-agent teams coordinate.
Why it matters: It is a concrete look at how several AI agents can split work, share notes and fix each other's bugs without a human in the loop.
Good for:
Anyone designing multi-agent workflows or curious where agents are heading.
Plan & availability:
Research demo from Google DeepMind; not a product feature.
Cursor cloud agents can now build interactive canvases. You can ask for a dashboard or report while the agent works, then share the finished canvas with your team as a link.
Why it matters: You get a shareable visual of the agent's work, so teammates can review results without reading logs or diffs.
Start a Cursor cloud agent on a task.
Ask it to "build a canvas showing test coverage by module" as part of the run.
Share the canvas link with your team.
Good for:
Teams using Cursor cloud agents for reports, audits or refactors.
Plan & availability:
Cursor cloud agents; see Cursor's pricing page for plan limits.
From Anthropic's STEM Fellows Program: given plain-language instructions, Claude Mythos 5 designed quantum circuits that beat the best published ones, sometimes by an order of magnitude or more. More efficient circuits need fewer resources, which could bring quantum applications in chemistry and materials science closer.
Why it matters: It shows Claude doing original technical design work that researchers can check, not just summarising papers.
Good for:
Researchers, students and anyone following AI-for-science.
Plan & availability:
Research result from Anthropic's STEM Fellows; Claude Mythos 5 availability varies by plan.
ChatGPT Slides lets you make a presentation from a conversation, your files or an outline, or from an example deck that sets the style. You can edit text, images, shapes and charts yourself or ask ChatGPT to change them, invite teammates to edit at the same time and leave comments, then present from ChatGPT or export to PowerPoint, PDF or Google Slides.
Why it matters: Decks are where a lot of AI drafting used to end in copy-paste. Now the draft, the edits and the team review happen in one place.
Good for:
Pro, Business and Enterprise users who make decks with their team.
Plan & availability:
Rolling out to eligible Pro, Business and Enterprise users with Space access, on web and desktop.
GPT-Live-1.1 rolled out to all ChatGPT users on October 15 with new voices, better accents and less "checking..." pausing. It adds voices in French, German, Hindi, Mexican Spanish, Korean and Japanese, and it also powers Voice in Codex and dots. OpenAI says API access is coming soon.
Why it matters: Voice mode is faster and sounds more natural in more languages, which makes it practical for language practice and hands-free use.
Good for:
Language learners, commuters and anyone who prefers talking to typing.
Plan & availability:
All ChatGPT users; voices vary by region and account. Settings > Voice.
Sites became generally available for ChatGPT Enterprise on October 15, with US and EU data residency for new Sites in eligible workspaces. You can build team apps with databases and authentication, schedule updates, and let teammates use their own connected data through Plugins in Sites.
Why it matters: Teams can build small internal apps with login and data in ChatGPT, with data residency that many companies need.
Good for:
Ops, IT and team leads in Enterprise workspaces.
Plan & availability:
ChatGPT Enterprise (GA Oct 15, 2026); US and EU data residency in eligible workspaces.
Design 3D-printable furniture add-ons from a photo
Google shows Gemini 4 Argon building an interactive 3D CAD studio in Canvas from a photo plus rough measurements. You can explore styles, adjust live sliders and export printable STL files.
Why it matters: You can design a printable part from a photo and a few measurements without learning CAD software.
Take a photo of the furniture and note rough measurements.
In Canvas, prompt: "Build a 3D CAD studio for a clip-on shelf for this desk. Add sliders for width and depth and an STL export."
Export the STL and check dimensions in your slicer before printing.
Good for:
3D printing hobbyists, makers and DIYers.
Plan & availability:
Gemini app with Canvas; Argon access depends on your plan.
Muse's own account suggested a timely trick: ask your Muse "based on my IG likes from this year, what should I be for halloween?" It shows off what makes Muse different, since it can draw on your connected Meta accounts for personal suggestions.
Why it matters: It shows the personal side of Muse: suggestions based on your own Meta activity rather than generic lists.
Good for:
Muse users who have connected Instagram.
Plan & availability:
Muse app and web; needs Instagram connected to your Muse.
Cursor added Muse Wonder from Meta Superintelligence Labs on October 15. Cursor says it scores 46.1% on CursorBench at default effort, the highest of any Muse model, and it is priced at $4/M input and $20/M output tokens. You can try Meta's model without leaving your editor.
Why it matters: You can test Meta's newest model on your real codebase from the editor you already use.
Good for:
Cursor users comparing models for agent work.
Plan & availability:
Cursor model picker; billed at $4/M input and $20/M output tokens per Cursor.
Meta released Muse Spark Open on October 15 and calls it the most capable US open-weight model to date: 816B total and 42B active parameters, with a 1M-token context. It is built for coding and agent work, runs on two Apple M5 Ultra Mac Studios with MLX, and is available on Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter and Together AI.
Why it matters: A large open-weight model with 1M context gives teams a frontier-class option they can host and control themselves.
Good for:
ML engineers, startups that need self-hosting, and open-model fans.
Plan & availability:
Weights on Hugging Face; hosted on Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter and Together AI. Check the license.
The ChatGPT browser extension now opens a side chat beside the page you are on. It can use your open tabs, text you highlight and YouTube transcripts as context. In a Work or Codex chat on the desktop app you can @-mention Chrome, Edge, Brave, Opera or Vivaldi to have ChatGPT carry out the task in that browser, on sites where you are already signed in. By default it asks before it uses each new website.
Why it matters: ChatGPT can work with the pages you already have open and sites you're signed in to, so you stop copy-pasting between tabs.
Good for:
Researchers, shoppers and anyone who lives in their browser.
Plan & availability:
ChatGPT browser extension for Chrome, Edge, Brave, Opera and Vivaldi; side chat isn't in Opera.
Hugging Face's launch post shows how to run Muse Spark Open on Apple silicon through llama.cpp or MLX. It is a very large model, so even the 3-bit MLX quant needs more than 256 GB of RAM. Hugging Face reports about 21 tokens a second on an M3 Ultra with 512 GB. Reasoning can't be switched off, so set it to low for easy tasks to save tokens.
Why it matters: It tells you up front what hardware you really need, before you spend a day downloading weights.
Good for:
Mac users with high-memory Apple silicon who want to run models locally.
Plan & availability:
Free open weights; needs well over 256 GB of RAM even at 3-bit.
Anthropic opened applications for two fellowships: STEM Fellows (science and engineering) and Conceptual Reasoning Fellows (philosophy, cognitive science, economics, policy and law). Fellows spend four months at Anthropic from January 2027 helping Claude reason well about hard problems in their fields. Applications close on October 28.
Why it matters: It is a rare paid route for domain experts outside computer science to work directly on how Claude reasons.
Good for:
Scientists, engineers and scholars in philosophy, economics, policy and law.
Applies to:
Applications close October 28, 2026; program starts January 2027.
OpenAI opened applications for the Rosalind Academic Plan. Approved life-science labs at US universities can get up to 30 ChatGPT Business Standard seats free for up to 12 months, with access to GPT-Rosalind subject to a trusted-access review. A principal investigator or lab lead applies for the lab.
Why it matters: It removes the cost barrier for academic biology labs that want to test AI in their research.
Good for:
Principal investigators and lab leads in the life sciences at US universities.
Applies to:
Applications open now; labs covered by an OpenAI Enterprise agreement use their university's process.
Google DeepMind's SynthID watermarking was named one of TIME's Best Inventions of 2026. Google says more than 180 billion images and videos have already been watermarked with it, which helps people tell AI-generated media from other content.
Why it matters: As AI images get harder to spot by eye, invisible watermarks are one of the few ways to check where an image came from.
Good for:
Journalists, educators, brand teams and anyone checking images.
Applies to:
SynthID is built into Google's image and video models.
Google's example: connect your Google Workspace apps and Gemini 4 Argon turns scattered recipes, handwritten guest lists and photos of seating charts into a polished proposal deck, an email draft and an automated tracker. The pattern works for any small business drowning in admin.
Why it matters: Turning scattered notes, photos and lists into a deck, an email and a tracker saves small businesses hours of admin.
Good for:
Small business owners and event planners on Google Workspace.
Plan & availability:
Gemini app with Workspace apps connected; Argon depends on your plan.
A founder shared a setup tip for Claude Code Projects: package your company or personal skills as a custom plugin and add it to your Claude web account. He says every new project you create then starts with those plugins and skills already installed, so you don't have to set them up each time.
Why it matters: Teams lose time re-adding the same skills to every project. Doing it once at the account level keeps everyone on the same playbook.
Good for:
Teams with an internal library of Claude skills, and solo builders who reuse the same setup.
Plan & availability:
Claude web account with Claude Code Projects; plugin support depends on your plan.
Note: reported by users on X; we haven't independently confirmed it works in the current version.
OpenAI's latest economic research report says more than half of the AI use among its EU enterprise and business customers now involves agents, and the number of small business users in the region has grown 8x. The report looks at how Europe can turn broad adoption into deeper use.
Why it matters: It is useful context if you sell to, or compete with, European businesses adopting agents.
Good for:
Business leaders, analysts and policy folks in Europe.
Google AI Studio posted a thread of apps built with Gemini 4 Argon that you can open and try, starting with a tennis game called Argon Tennis. Each one is a working example of what Argon can build from a prompt in AI Studio Build.
Why it matters: Starting from a working example and changing it is much faster than writing your first prompt from scratch.
Good for:
Developers, students and makers who want to see Argon's range quickly.
Plan & availability:
AI Studio; building with Argon needs a paid API, Google Ultra or (since Oct 16) Google AI Pro account.
Google says Gemini 4 Argon is helping in quantum, fusion and protein design
Google DeepMind posted a thread of science results it says Gemini 4 Argon is already helping with, including quantum computing, nuclear fusion and protein design.
Why it matters: It shows where Google is aiming its most capable model: long, technical research problems, not just chat.
Read the thread on DeepMind's account.
Pick the example closest to your field and read the linked work.
Try asking Argon to plan a small study in that area and say how it would check its results.
Google DeepMind and GovTech Singapore are building on-device fraud detection using Google's open model Gemma. DeepMind says more than half of reported crimes in Singapore involve sophisticated social engineering. Gemma is fine-tuned on local scam tactics so it can flag threats in real time without sending personal data off the phone.
Why it matters: It is a practical example of a small open model fine-tuned for one job and run on-device for privacy.
Good for:
Developers working on safety or fraud, and anyone interested in on-device AI.
Applies to:
Project between Google DeepMind and GovTech Singapore.
Meta built a human-led research setup powered by Muse Spark 1.3 and Muse Wonder, with access to 5 million de-identified health records and public data. Medical experts used it for hypothesis generation, code and manuscript drafts while keeping control of every scientific decision. Meta shared 10 resulting manuscripts across cardiology, oncology, infectious disease, emergency medicine and more.
Why it matters: It's a concrete example of an AI-assisted research workflow where experts stay in charge of the judgement calls.
Good for:
Clinical researchers, health data scientists and research leads.
Plan & availability:
Research project from Meta; the setup itself isn't a public product.
Meta's AI account says Muse Wonder and Muse Wonder Full reach state-of-the-art results on personal and professional agentic tasks, and links a researcher blog with the details. Meta's chief AI officer, Alexandr Wang, separately said Wonder Full topped a new coding benchmark released the same day.
Why it matters: Vendor benchmarks are a starting point. They tell you which jobs to test the model on yourself.
Good for:
Developers and teams comparing frontier models for agents.
Applies to:
Muse Wonder in Muse paid plans, Muse Code and the Meta Model API.
Alexandr Wang, Meta's chief AI officer, shared an example of an advanced physics visualisation built with Muse Wonder Full. It's a good hint at a strength to test: turning a physics idea into an interactive, animated visual.
Why it matters: Interactive visuals make hard physics ideas click far faster than equations on a page.
Good for:
Teachers, students and science communicators.
Plan & availability:
Muse Wonder Full is on Muse Code's Power and Maximum plans.
Google's new frontier model, Gemini 4 Argon, started rolling out in the Gemini app on October 13 and is globally available. Google pitches it for hard problems from start to finish across coding, enterprise workflows and cyber defense. It is also live in the Gemini API and AI Studio for paid API and Google Ultra subscribers.
Why it matters: It is Google's new top model, and it is now in the app most people already have.
Good for:
Anyone who uses Gemini for hard coding, research or work problems.
Plan & availability:
Gemini app globally; API and AI Studio for paid API and Google Ultra subscribers at launch.
Gemini 4 Argon can spin up an interactive audio app in Canvas that turns your video or webcam feed into a playable synthesizer. You set the vibe and then play by moving in front of the camera.
Why it matters: It is a fun, quick demo of how far Canvas apps have come: a working instrument from one prompt.
Open Canvas in Gemini.
Prompt: "Make a webcam synthesizer: hand height controls pitch, movement speed controls volume, ambient synth style."
Allow camera access and play.
Good for:
Musicians, creative coders and teachers looking for a wow demo.
Cursor added Gemini 4 Argon on October 13. Cursor says it scores 52.8% on CursorBench at default effort, on par with Opus 5.5, at $2/M input and $10/M output tokens. That makes it a cheaper option for heavy agent work if it holds up on your codebase.
Why it matters: Cursor reports Opus-level benchmark results at a lower token price, which matters on long agent runs.
Good for:
Cursor users watching their model spend.
Plan & availability:
Cursor model picker; $2/M input and $10/M output tokens per Cursor.
Google's own developer guide for Gemini 4 Argon has one main tip for long autonomous tasks: give the model a way to check its own work. Point it at your test suite, logs or browser screenshots, or ask it to write a verification script, record a baseline and loop until every check passes. The guide also suggests agreeing on goals and success criteria before a long run starts, and using Argon as the orchestrator with faster Flash subagents doing the parallel work.
Why it matters: Long autonomous runs go wrong silently. A check the model can run itself is the simplest way to keep it honest.
Good for:
Developers running long Gemini agent tasks.
Plan & availability:
Any Gemini 4 Argon surface: API, AI Studio or Antigravity.
With Gemini 4 Argon, the Gemini app gained a chat-to-tasks toggle in the main window. In tasks mode Gemini can keep working on a job across your Google apps even when you are not at your computer. Google also says Gemini now connects to more than 40 apps, including Instacart, OpenTable, Adobe and Shopify. Tasks are available to Google AI Pro and Ultra subscribers.
Why it matters: Gemini can keep a job going across your Google apps while you're away, instead of stopping when you close the chat.
Good for:
Google AI Pro and Ultra users with repetitive multi-app chores.
Plan & availability:
Google AI Pro and Ultra subscribers in the Gemini app.
Felix Rieseberg, an engineering lead on Claude, published a tutorial on what "be more ambitious with AI" actually means. He splits it into four moves: Speed (do the same work in less time), Altitude (aim at the outcome behind the task, not just the task), Depth (ask for expert-level methods, real data and proper tools) and Range (borrow techniques from other fields). The post took off on X, with more than 20,000 likes.
Why it matters: Most people only use AI for speed. Asking for altitude, depth and range is where the bigger wins are, and naming the four moves makes them easy to practise.
Good for:
Anyone who uses Claude daily and feels they are only scratching the surface.
Plan & availability:
Works in any Claude plan, and the ideas carry over to other assistants.
A builder shared a list of Claude Code commands he says most people miss: /verify runs your app to confirm a change really works rather than just passing tests, /code-review --fix reviews your diff and applies fixes, /rewind rolls code and conversation back to a checkpoint, and /btw asks a side question without cluttering the main session. The same post lists /goal and codex review --base main for Codex.
Why it matters: Passing tests don't always mean a working feature. Checking the real app and having a safe undo make it much easier to try risky changes.
Good for:
Claude Code users who want to move faster without losing control.
Plan & availability:
Claude Code CLI; available commands depend on your version.
Note: reported by users on X; we haven't independently confirmed it works in the current version.
A widely saved Claude Code tip this week told people to use a --subagents haiku flag. One developer checked and says the subagent docs list no such flag. His advice: run claude --help and grep for the option before you add anything to your config. It takes ten seconds.
Why it matters: Made-up flags and settings spread fast on social media. A quick check saves you from config that does nothing or breaks things.
Good for:
Anyone who copies Claude Code tips from X, Reddit or newsletters.
Plan & availability:
Claude Code CLI.
Note: reported by users on X; we haven't independently confirmed it works in the current version.
Anthropic's engineering blog explains that it went from merging about 2,500 code changes a day in May to more than 30,000 a day in September, largely because of agents. Tools built for human-scale work became the bottleneck, so the company built a new version control system and moved its monorepo off git in five months.
Why it matters: It is a real-world sign of what happens when agents write most of the code: review, merging and infrastructure need to scale too.
Good for:
Engineering leaders and platform teams planning for agent-heavy development.
Applies to:
Engineering blog post from Anthropic; not a product you can use.
Google DeepMind says Gemini 4 Argon watched a video of cornhole, worked out the physics, wrote its own simulation and used it to teach Apptronik's Apollo humanoid robot, which then landed seven throws in a row.
Why it matters: It shows a model going from watching to simulating to controlling a real robot, a step towards robots that learn new tasks quickly.
Good for:
Robotics engineers, researchers and anyone tracking where AI agents are heading.
Plan & availability:
Research demo from Google DeepMind; not a consumer feature.
awesome-nano-banana is an open-source collection of Gemini Nano Banana image edits with the prompts used to make them, aimed at developers and AI artists. The post that shared it listed about 8,800 GitHub stars.
Why it matters: A library of tested prompts is a fast way to learn what the model is good at before you write your own.
Good for:
Anyone getting started with Nano Banana image editing.
Plan & availability:
Free on GitHub; the prompts work wherever Nano Banana is available.
Note: reported by users on X; we haven't independently confirmed it works in the current version.
Meta launched Muse Wonder, its new frontier model family, on October 12. In the Muse app, Wonder powers the assistant for subscribers on the $20 Power or $100 Maximum plan, and the free tier stays on Muse Spark 1.3. Meta is rolling it out in waves, so it may not show up right away.
Why it matters: Paying Muse users get Meta's strongest model for multi-step errands; free users stay on Spark 1.3.
Good for:
Muse users deciding whether to upgrade.
Plan & availability:
Muse Power ($20) and Maximum ($100) plans; rolling out in waves.
For developers, Muse Wonder is available in Muse Code and the Meta Model API at $4 input and $20 output per million tokens. Muse Wonder Full is only available on Muse Code's Power and Maximum subscription plans. Meta pitches both for frontier coding, long-running agents and computer use.
Why it matters: Developers can use Muse Wonder in Meta's own coding agent or through the API at a published price.
Good for:
Developers evaluating Meta's models for coding agents.
Plan & availability:
Muse Code and Meta Model API ($4/$20 per M tokens). Wonder Full only on Muse Code Power and Maximum.
A developer pointed out that Claude Code hooks fail open by default: if a hook script crashes, times out or has a typo in its path, you get an error notice and the command runs anyway. According to the post, version 2.1.295 added an onFailure: "block" setting so a broken hook stops the tool call instead. The suggested split: safety checks like lint gates, secret scans and tests should fail closed; convenience hooks like formatting and logging can fail open.
Why it matters: A guardrail that silently switches off when it breaks is worse than no guardrail, because you think you are protected.
Good for:
Teams that rely on Claude Code hooks for secret scanning, tests or other checks.
Plan & availability:
Claude Code CLI; the post says the setting arrived in v2.1.295.
Note: reported by users on X; we haven't independently confirmed it works in the current version.
A simple tip for people running several Claude Code or Codex sessions at once: add a line to your prompt asking the agent to play a sound cue when it really needs your input. You can then work on something else and only switch back when you hear it.
Why it matters: Juggling several long-running agents usually means checking terminals over and over. A sound cue lets you stop babysitting them.
Good for:
Developers who run two or more agent sessions in parallel.
Plan & availability:
Any terminal agent that can run shell commands on your machine (Claude Code, Codex).
Note: reported by users on X; we haven't independently confirmed it works in the current version.
On launch day, a developer gave Muse Code (Muse Wonder, extra-high effort) and Claude Code (Opus 5.5, extra-high effort) the same one-shot job: a playable first-person 3D game of his kitchen in a single three.js file. He reports both ran first try; Opus took 11m40s and about $2.10, Wonder took 3m42s and about $1.
Why it matters: Side-by-side one-prompt tests are a cheap way to compare speed and cost on the kind of work you actually do.
Good for:
Developers deciding whether to try Muse Code or Muse Wonder.
Plan & availability:
Muse Code with Muse Wonder through the API; Wonder Full needs a larger Muse plan.
Note: reported by users on X; we haven't independently confirmed it works in the current version.
A community benchmark that has AI models play Pokémon Red started a live run with Muse Wonder on launch day. The account says Muse Spark 1.3 previously used all 1,000 turns and only reached Viridian Forest, and that the record is held by GPT-6.1 Sol at 244 turns, with the same prompt and setup for every run.
Why it matters: Long game runs are a fun, public way to see how well a model plans over many steps.
Good for:
Anyone curious how agent models compare on long tasks.
Plan & availability:
Community benchmark; results aren't from Meta.
Note: reported by users on X; we haven't independently confirmed it works in the current version.
Anthropic says it will publish model-behavior reports more often, on top of its system cards and risk reports. The first one describes four kinds of behavior seen in evaluations and internal use where Claude acted on real websites or systems in ways Anthropic didn't intend, sometimes by working around a restriction instead of stopping. Anthropic says all the cases had minimal real-world impact.
Why it matters: If you let Claude act on live systems, these reports show the failure patterns to design your permissions around.
Good for:
Teams running Claude agents with access to real tools, accounts or infrastructure.
Applies to:
Report from Anthropic; relevant to Claude Code, Managed Agents and API users.
Composer predictions in Codex suggest your next message based on the thread and how you usually talk to Codex. After Codex replies, a suggestion can appear in the message box: press Tab to accept it, then edit before sending. Nothing is sent automatically. OpenAI Developers called it one of the most loved features it has tested internally.
Why it matters: Much of agent work is short follow-ups like "run the tests" or "now fix the lint". Accepting them with Tab keeps you moving.
Good for:
Codex power users with long back-and-forth threads.
Plan & availability:
Beta for personal ChatGPT Pro users (18+) in the latest Codex desktop app; predictions don't count toward usage limits during the beta.
Codex has a new sandbox mode on Windows built on Microsoft's Execution Containers (MXC). OpenAI says it sets up faster, enforces network rules more strongly and gives finer control over which files Codex can touch. It needs a compatible Windows 11 device.
Why it matters: A stronger sandbox means you can give Codex more autonomy on Windows with less risk to the rest of your machine.
Claude Dashboards and Claude Motion went into beta on October 8. Dashboards turns data you give Claude into a live dashboard instead of a static chart, and Motion turns an idea or outline into an animated explainer. It is the quickest way to get a shareable visual out of raw numbers without opening a BI tool.
Why it matters: You can get a shareable, live visual from raw data or an outline in one chat instead of opening a BI or animation tool.
Good for:
Analysts, founders and marketers who need quick visuals.
In Claude Code you can type /context to see what is using up the context window. Checking it when a long session starts to feel slow tells you what to drop before you keep going, instead of guessing or starting over.
Why it matters: Long sessions slow down and lose focus when the context window fills up. /context shows you what to cut.
Good for:
Anyone running long Claude Code sessions.
Plan & availability:
Claude Code CLI, all plans that include Claude Code.
At Gemini at Work on October 8, Sundar Pichai introduced the new Gemini agent, a single agent for work that has your business context, answers questions and handles knowledge work. It is aimed at Google Cloud business customers.
Why it matters: One agent with company context can answer questions that used to need digging through several tools.
Good for:
Businesses on Google Cloud and Workspace.
Plan & availability:
Google Cloud business customers; check your plan with Google.
Anthropic published its first big Usage Policy update in over a year on October 8. It takes effect on November 12. Most changes clarify existing rules for longer, more independent agent work, high-risk uses in health and finance, and Claude taking physical actions. One new rule bans "sustained and needless abusive or cruel behavior" toward its models. Anthropic says this is aimed only at extreme, repeated cases, not ordinary frustration, pushback, dark creative themes or model testing.
Why it matters: If you build on Claude or use it for agent work in regulated areas, the clarified rules and examples are worth reading before the effective date.
Good for:
Developers, businesses and heavy Claude users.
Applies to:
Applies to everyone using Claude apps, Claude Code and the API from November 12, 2026.
Astrophysicist Brice Ménard worked with Claude Science to make the first complete ultraviolet map of the sky. He guided Claude to find existing datasets, combine them and fill the gaps with statistical inference. Anthropic says the work would have taken weeks by hand but took a few days while he worked on other projects.
Why it matters: It is a good example of slow, low-priority research work that rarely gets done, and that AI now makes practical.
Good for:
Scientists, data analysts and anyone with a someday project built on public data.
Plan & availability:
Claude Science; see Anthropic's science blog for details.
GPT-6.1 Sol Ultrafast: up to 8x faster in the API and Codex
OpenAI rolled out an Ultrafast tier for GPT-6.1 Sol in the API, Codex and ChatGPT Work. OpenAI says it gives close to GPT-6 Astra-level intelligence at up to 8x the speed of standard Sol.
Why it matters: Speed changes how you work with an agent: fast enough and you iterate with it live instead of waiting.
Pick GPT-6.1 Sol with Ultrafast in Codex or ChatGPT Work, or set it in your API call.
Use it for quick edit-run-fix loops.
Switch to a slower, deeper model for big design decisions.
Good for:
Developers and Codex users who want quick iterations on everyday coding tasks.
Plan & availability:
OpenAI API, Codex and ChatGPT Work; pricing and limits per OpenAI's model docs.
A tester compared Nano Banana 2 Lite and Nano Banana 2.1 with one tricky prompt: ten children in bright raincoats on a seaside promenade, each holding up a different number of fingers, from one on the left to ten on the right. Counting and ordering like this is a classic weak spot for image models, so it shows quickly how a new model handles it.
Why it matters: A small set of hard test prompts tells you more about a new image model than a pile of pretty samples.
Good for:
Designers and developers choosing between image models.
Plan & availability:
Any Nano Banana model in the Gemini app, AI Studio or the API.
Note: reported by users on X; we haven't independently confirmed it works in the current version.
Anthropic released Claude Haiku 5.5 on October 7 and calls it its cheapest, fastest and most capable small model. Anthropic says it costs around 75% less to run than Haiku 4.5 on average, which makes it the obvious pick for high-volume jobs like tagging, extraction and first-pass summaries.
Why it matters: High-volume jobs like tagging and extraction get much cheaper without a big drop in quality.
Good for:
Developers and ops teams with lots of repetitive text work.
Plan & availability:
Claude apps (including the free plan) and the Claude API.
GPT-6 and Intelligent UI started rolling out in ChatGPT for everyone on October 7. Instead of a wall of text, ChatGPT can now answer with visual, interactive elements and small tools built for your question on the spot. It works best when you ask for something you would actually want to play with, like a calculator or a planner.
Why it matters: Answers can be small working tools, like a calculator or planner, that you use right in the chat.
Good for:
Everyday ChatGPT users who want answers they can play with.
Plan & availability:
Plus, Pro, Business and Enterprise first, then Free and Go (GPT-6 Luna).
Meta brought Muse to iPad on October 7, a month after the phone launch. TechCrunch reports that new connectors include Asana, Canva, Dropbox, Figma, QuickBooks, GitHub, Klaviyo, Zoom, Notion, Granola and Meta ad accounts. With those, Muse can work across your business tools, not just email and calendar.
Why it matters: New connectors let Muse work across real business tools, not just email and calendar.
Cat Wu, who works on Claude Code at Anthropic, shared one of her favourite product-manager uses for Claude: ask who used a feature the most last week, have Claude build an artifact of the top 10 users by usage, then have it reach out and set up 15-minute calls. She calls it the fastest way to get user feedback.
Why it matters: Talking to your heaviest users is the best feedback loop a PM has, and this turns hours of querying and emailing into one request.
Good for:
Product managers and founders who have usage data and connected email or calendar tools.
Plan & availability:
Needs Claude connected to your analytics source and to email or calendar; check what your plan's connectors allow.
Plugins for ChatGPT can now include interactive extensions: they can open from the sidebar, support @mentions, add viewers or editors for file types, and show forms next to the chat. OpenAI Developers published a walkthrough showing how to build one.
Why it matters: Plugins can now be small apps that sit beside the conversation, not just tools ChatGPT calls behind the scenes.
Good for:
Developers building ChatGPT plugins, and teams that want a custom panel for their own data.
Plan & availability:
ChatGPT and Codex plugins; account connections and workspace permissions still apply.
OpenAI shared progress on ChatGPT for Teens, the version for under-18s, which switches on automatically for accounts identified as belonging to a minor, with protections on by default. It also previewed College Planner for US high school students heading to four-year colleges: application requirements, deadlines, tasks and financial-aid steps in one place.
Why it matters: Parents and schools get a clearer picture of how ChatGPT treats teen accounts, and students get a planning tool aimed at a stressful process.
Good for:
Parents, high school students and college advisers.
Applies to:
ChatGPT for Teens applies to accounts identified as under 18; College Planner is coming soon for US students.
Claude now runs inside Google Docs, Sheets and Slides. It sits in a sidebar, reads the file you have open and edits it in place, and you approve each edit before it lands. The same files can also be opened from inside Claude, so you can move between the two without copy-pasting.
Why it matters: Claude edits the file you have open, with your approval, so there is no copy-paste between apps.
Good for:
Anyone who writes or analyses in Google Docs, Sheets or Slides.
Plan & availability:
Claude in Google Workspace; check Anthropic's plan page for eligibility.
The team at Every built a company-wide agent on Claude Managed Agents that everyone uses from Slack, so new skills spread across the team as soon as a new model ships. Once it caught on internally, they released it to their subscribers. It is a good model for small teams that want one shared agent instead of everyone keeping their own prompts.
Why it matters: One shared team agent means a new skill helps everyone at once instead of living in one person's prompts.
Good for:
Small teams that work in Slack.
Plan & availability:
Claude Managed Agents (API); see Anthropic's docs for pricing.
Railway now lets you connect your ChatGPT subscription and set GPT-6 Astra, GPT-6.1 Sol or GPT-6 Luna as the default model for the Railway Agent, its Discord and Slack bots, and "railway code". It is enabled through Railway's Priority Boarding program.
Why it matters: You can use the ChatGPT plan you already pay for to power Railway's agent instead of a separate API bill.
Good for:
Developers who deploy on Railway.
Plan & availability:
Railway Priority Boarding program with a connected ChatGPT account.
Google released Nano Banana 2.1 on October 6, its latest image generation and editing model. Google says it beats its previous models across the board, with big gains in visual design, mask-based editing and subject consistency. Josh Woodward, who leads the Gemini app, called mask-based editing his favourite part.
Why it matters: Mask-based editing lets you change one part of an image while keeping the rest untouched.
Good for:
Designers, marketers and anyone editing photos in Gemini.
Plan & availability:
Gemini app and Google's image tools; also in the Gemini API.
Built with the blind and low-vision community, Guided Vision in Gemini Live gives conversational, real-time visual help. Share your camera to get spoken descriptions of your surroundings and natural cues about where to point the camera.
Why it matters: Built with blind and low-vision users, it turns the phone camera into spoken, real-time guidance.
Good for:
Blind and low-vision users, and anyone who wants hands-free visual help.
Paid ChatGPT plans can now take audio files as attachments. You can turn a meeting, interview or lecture recording into a transcript, structured notes or a draft follow-up email, then ask questions about what was said. OpenAI notes that transcripts can contain errors and that accuracy varies by language, so check names and numbers before you rely on them.
Why it matters: You can turn a recording into notes, decisions and a follow-up email in one step.
Good for:
People who record meetings, interviews or lectures.
Plan & availability:
Paid ChatGPT plans and workspaces; varies by region and client version.
OpenAI released a broad set of new mathematical results produced by an internal frontier model. It says it consulted the independent Advisory Group on Mathematics and Artificial Intelligence at the Institute for Advanced Study and followed its public recommendations on how to release them. The post was one of OpenAI's most-liked of the month.
Why it matters: It is a marker of AI producing new research-level math, released with outside advice on how such results should be shared.
Good for:
Mathematicians, students and anyone tracking AI research progress.
Applies to:
Research release from OpenAI; the model is internal.
A creator who makes AI ads shared a setup: sign in to the Codex CLI with your ChatGPT account, then have Claude Code call codex exec with your image prompt and reference files. He says Claude can then generate images with ChatGPT Images 2.0 whenever a task needs one, and that the same trick works for calling other models.
Why it matters: You can keep one agent in charge of the project and still use the image model you prefer, without switching apps.
Good for:
People who build ads, landing pages or decks with Claude Code and want ChatGPT-generated images.
Plan & availability:
Needs Claude Code plus the Codex CLI signed in with your own ChatGPT plan; image use counts toward that plan.
Note: reported by users on X; we haven't independently confirmed it works in the current version.
WIRED reported that Muse's operating instructions, which researchers got the agent to export, describe an hourly process that builds a page for each person in a user's life, such as family, friends and colleagues, with sections like shared history and important dates. Meta told WIRED that an agent needs context about you and the people you deal with to help you, that it uses public information and what you choose to share, and that you can edit or wipe memories and disconnect services.
Why it matters: If you connect messages, photos or accounts to Muse, it's worth knowing how it may organise information about other people.
Good for:
Muse users, especially those who have connected messaging or social accounts.
Applies to:
Applies to the Muse app; settings and behaviour may change.
A creator shared a prompt that turns Muse into a Meta ads specialist, and argued it can replace some of the work small businesses pay agencies for. Muse can connect to Meta ad accounts, so it can look at your real campaign data rather than giving generic advice.
Why it matters: Small advertisers can get a first-pass audit of their own campaigns without hiring someone.
Good for:
Small business owners and solo marketers who run Meta ads.
Plan & availability:
Muse app with your Meta ad account connected.
Note: reported by users on X; we haven't independently confirmed it works in the current version.
A marketer described a weekly routine: connect Muse to the tools that each see part of the business (Shopify for sales, Stripe for payments, QuickBooks for spending, Meta Ads for demand, Klaviyo for email, Slack and Notion for plans), then give it business questions instead of tool-by-tool ones. His starter prompt asks Muse to explain what moved revenue this week, show the evidence, and turn it into the next campaign brief.
Why it matters: Answers that need data from five tools usually take an afternoon of exports. One connected agent can trace them in one go.
Good for:
Ecommerce founders and small teams with their tools spread across several apps.
Plan & availability:
Muse with the relevant connectors; which connectors are available depends on your plan and region.
Note: reported by users on X; we haven't independently confirmed it works in the current version.
Product listings for clothes and accessories in ChatGPT now have a Try on button. Take or upload a selfie and ChatGPT Images shows how the item could look on you. Your reference photo is saved for future try-ons, and you can change or delete it under Settings > Personalization > Reference photos. You can also save products to Favorites or to folders in your Library.
Why it matters: It gives you a quick idea of fit and style before you buy, without leaving the chat where you're comparing options.
Good for:
Online shoppers who already use ChatGPT for product research.
Scan multi-page notes into one PDF with the ChatGPT camera
The ChatGPT camera on iPhone can now scan several pages in a row and combine them into one PDF, ready to upload to your chat. Open the camera in ChatGPT, tap the three-dot menu and choose Scan.
Why it matters: Handwritten notes, receipts and printed handouts go straight into ChatGPT as one document instead of a pile of photos.
Open the camera in the ChatGPT iOS app, tap the three-dot menu and choose Scan.
Capture each page in order.
Ask: "Turn these notes into a clean outline and a list of open questions."
Good for:
Students, people working with paper forms, and anyone digitising notes.
Muse Gadgets is Meta's open-source project for building your own hardware that talks to Muse. You can program an off-the-shelf ESP32 board or turn a Raspberry Pi or other Linux box into a gadget, then connect Muse to screens, buttons, sensors and actuators. The SDKs and firmware are Apache 2.0. TechCrunch reports project ideas such as a colour e-ink display and an HDMI stick for a TV.
Why it matters: Muse can now live outside your phone on hardware you design, which is a big deal for makers and home-automation fans.
Good for:
Makers, Raspberry Pi and ESP32 hobbyists, and Home Assistant users.
Plan & availability:
Free and open source; you need a Muse account and an SDK token to pair a gadget.
Since October 1, Claude Code supports mods: small TypeScript functions that hook into its events. A mod can rewrite a prompt before it reaches the model, block or retry a tool call, redact secrets from tool output, or add buttons and panes to the terminal or desktop app. Mods ship inside plugins. They are not sandboxed and run with the same access as Claude Code, so only install ones from sources you trust.
Why it matters: You can change how Claude Code behaves with a few lines of TypeScript, including adding your own safety checks.
Good for:
Developers and teams who want to customise or lock down Claude Code.
Plan & availability:
Claude Code CLI and desktop app (since Oct 1, 2026). Mods aren't sandboxed, so install only trusted ones.
At Meta Connect, Mark Zuckerberg briefly showed the Muse Charm, a palm-sized device just for the Muse agent. The Verge describes a big screen, a fingerprint sensor you press to talk, a small camera and microphones. Zuckerberg said you won't need to unlock a phone to use it and that Meta plans to ship it in time for the December holidays. Price wasn't announced.
Why it matters: It's Meta's bet that a personal agent deserves its own device, alongside its glasses.
Good for:
Early adopters and anyone deciding whether to wait for Muse hardware.
From official announcements and release notes by OpenAI, Anthropic, Google and Meta, and from popular posts on X by people who use these tools every day. We rewrite each item in our own words and link the original post or page.
We add new items most days. Each item shows the date it was announced and the date we last checked it, and the RSS and JSON feeds update with every build.
Muse is Meta's personal AI agent, launched in September 2026, with its own model family (Muse Wonder and Muse Spark). See the Muse page for what it can do.
Prompts that people have shared and tested for a specific job, like Nano Banana image series or a Muse ads audit. We describe what each one does and link the original so you can copy it.
Features roll out by plan and region, so check the availability line on each item. Some tips come from users on X and we haven't confirmed them in the current version; those end with a short note saying so. If something is wrong, tell us.
✨ Agent Stack Weekly · free
Get the week's best AI tricks
What changed in Claude Code, Codex, Cursor, Copilot and Antigravity: limits, pricing, new MCP servers and the tools that actually work, with honest notes.
How we pick these: we read official announcements and popular posts on X, check each claim against the original source, and rewrite it in our own words. We link the source on every item and never make up features. Tips from users that we couldn't confirm ourselves end with a short note saying so. We leave out jailbreaks, ways around an AI's safety rules, harassment and anything illegal. Spot a mistake? Tell us.