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 · 13 STORIES

AI news, 9 November 2026

From the community

OpenAI starts selling a Codex-made Game Boy-style handheld for $200

OpenAI's Supply Co. store has listed the ModRetro Chromatic: Codex Edition at $200. It is a handheld that plays Game Boy and Game Boy Color cartridges, and OpenAI's page says you can use Codex and a ModRetro Chromatic plugin to build a game in GB Studio, test it in a browser emulator, and write it to the included programmable Codex Cartridge. It ships only to the US, Canada, Australia, New Zealand, Japan, the UK and the EU. At the time of checking the page said Restock soon, with an expected restock date of 19 November, so it may not be buyable right now. A matching Chromatic Case is $35. It follows the DevDay edition handed out on 29 September, and the teaser OpenAI Developers posted on 8 November.

Why it mattersThis is physical merchandise rather than a model launch, but it is a concrete, low-cost way to see what Codex can do beyond code: describe a game, get a ROM, and play it on real hardware. We have not ordered or tested it, so how well the Codex workflow works in practice is unconfirmed. OpenAI's page says to update the firmware and activate Developer-Mode before playing from a cartridge.

Sourcesopenai.com ↗ · openai.com ↗ · x.com post ↗

Anthropic launches Claude Fable 5.2 and Claude Mythos 5.2

Anthropic announced Claude Fable 5.2 and Claude Mythos 5.2 on 9 November. They are the same underlying model: Fable has stronger cybersecurity and biology safeguards and is generally available; Mythos is for vetted organisations via Anthropic's Life Sciences and Cyber Verification Programs. Anthropic says they improve coding, reasoning and vision, communicate more clearly than earlier Fable/Mythos models, and generate outputs up to 30% faster than 5.1. API pricing is $8 per million input tokens and $40 per million output tokens (down from $10 / $50 on 5.1), with an optional fast mode at $16 / $80. Both have a 1M-token context window and up to 128K output tokens. Model IDs: claude-fable-5-2 and claude-mythos-5-2. Anthropic says Fable 5.2 is available on paid Claude plans, the Claude Platform, AWS, Google Cloud and Microsoft Azure. On its own benchmarks it places the models a step above Opus 5.5 on most capabilities it evaluated; on FrontierCode 2.0 it says High effort scores about level with Opus 5.5 at Max for roughly half the cost per task. For most daily coding and knowledge work Anthropic still recommends starting with Opus 5.5.

Why it mattersIf you use Claude for long, hard technical projects, this is the new top tier, at a lower API price than 5.1. For everyday work Anthropic itself still points you at Opus 5.5 first. Mythos access needs a verification programme.

Sourcesx.com post ↗ · anthropic.com ↗ · platform.claude.com ↗ · platform.claude.com ↗ · platform.claude.com ↗Fact Hype or Fact? check

Google AI Studio can now build apps with Google Maps Platform

Google announced on 9 November that Google Maps Platform is available inside Google AI Studio. You describe a maps idea and AI Studio sets up Maps for you, with map tiles, Places, weather, routing, 3D buildings and Street View. Prototyping uses a Maps Demo Key: no credit card is required, and the daily limit is 100 calls per API. For production you link a Google Cloud billing account and use standard Maps Platform pricing, which includes monthly free usage per SKU. Try it at ai.studio/build. Google AI Studio is available in the UK and EU.

Why it mattersIf a store locator or route planner has been on your list, you can prototype one in AI Studio without setting up billing first. The demo quota is small (100 calls per day per API), so a live app still needs a billed Cloud project and the usual Maps Platform rates.

Sourcesx.com post ↗ · developers.googleblog.com ↗ · developers.google.com ↗ · developers.google.com ↗

DeepSeek Harness says it has 1.54 million daily users and 35,000 plugins

DeepSeek published an update on its agent app, DeepSeek Harness. It says it has 1.54 million daily users, about 35,000 community-built plugins and that around 60% of active users have installed or built one. It plans an official plugin marketplace, better plugin security and a support programme for plugin authors with token grants. Examples include a creator workbench, a design plugin and a paper-reading tool.

Why it mattersPlugins are how an agent gets taught your specific tasks, so a marketplace would make it easier to find ready-made workflows. All the numbers are DeepSeek's own claims and unverified, and plugins vary in quality, so check what a plugin does before installing it.

Sourcesx.com post ↗

Hermes Agent can now review your code for security holes

Nous Research says Hermes Agent has a new official Security plugin. It works out what an attacker could reach, splits the code across parallel helper agents, checks each suspected problem against the code, and can write and run a small test in a temporary copy of the project. The report lists the evidence and what was ruled out, and exports in a standard format other security tools read.

Why it mattersIf you run software, or pay someone to, this is a cheap first pass at finding security bugs before a human reviewer does. It is a first pass, not a replacement for a proper security review. Details are Nous Research's own description.

Sourcesx.com post ↗

Apple's Xcode plugin lands in Claude Code, and Greg Isenberg's playbook for using it

An Apple-authored "xcode" plugin was merged into Anthropic's official Claude Code plugin marketplace on 6 November. You install it with "/plugin install xcode@claude-plugins-official". It needs a Mac with a current Xcode, and it lets the agent build, run and test iOS apps, use SwiftUI previews and simulators, and read logs. Apple's developer documentation also explains how to connect Claude Code or Codex to Xcode's tools once you switch on "Allow External Agents to Use Xcode Tools" in Xcode's settings. Apple hasn't made a public announcement. @gregisenberg used the news for a playbook: sort utility apps by revenue in Appfigures or Sensor Tower, look for ones rated under 3.5 that haven't been updated in a while, then build a better version with agents. The playbook is his opinion, not a guarantee of revenue. Fact or hype: Fact. The plugin is in the official marketplace and Apple's docs describe the agent connection; there's just no Apple press release.

Why it mattersIf you've wanted an iPhone app but couldn't justify a developer, an agent can now drive Xcode for you. It still needs a Mac, an Apple developer account and someone to test the app properly before it goes to the App Store.

Sourcesx.com post ↗ · developer.apple.com ↗Fact Hype or Fact? check

What each ChatGPT Pro tier buys, and a 'Pro 50' that's only an idea

OpenAI's Codex lead, @thsottiaux, clarified that none of the ChatGPT Pro plans has a 5-hour limit, and that Pro 100, Pro 200 and Pro 500 give 10x, 20x and 50x the usage of Plus. Pro 500 also gets "ultrafast" access. OpenAI's help centre confirms the tiers. In a follow-up seen 1.3 million times, he asked whether a $50 "Pro 50" tier would make sense. That's a question, not a product: nothing has been announced. Fact or hype: Fact for the tiers (OpenAI's help page). The $50 tier is only being floated.

Why it mattersIf your team uses Codex or ChatGPT for agent work, match the tier to how many hours a day you actually run it. Don't plan a budget around a cheaper Pro 50 until OpenAI launches one.

Sourcesx.com post ↗ · help.openai.com ↗Fact Hype or Fact? check

'GPT-6.2 Sol' is still a rumour, but OpenAI posted 'Tomorrow.'

Over the weekend, many ChatGPT users said GPT-6.1 Sol seemed to be quietly routing them to a newer model, and people assumed it was "GPT-6.2 Sol". @kimmonismus and @Mr_Salio shared impressive-looking Three.js demos. On 8 November, @OpenAIDevs posted a 16-second teaser captioned "Tomorrow." (466,000 views). OpenAI hasn't announced GPT-6.2 Sol: its changelog still lists GPT-6.1 Sol (29 September) as the latest. Other weekend "drops" (Qwen 4, a "Hyperion" Gemini family spotted in Devin CLI, Gemini 4 Carbon) are also unconfirmed or not usable yet. Rumour: unconfirmed. Something from OpenAI looks likely today, but nobody outside OpenAI knows what it is.

Why it mattersDon't switch tools or plans on the strength of demo clips. Wait for the official release notes and test it on one of your own tasks first.

Sourcesx.com post ↗ · x.com post ↗

'Anthropic's progress per R&D dollar fell 63%': that's someone else's maths

@kimmonismus shared an analysis saying Anthropic's measured progress nearly doubled while its R&D spending rose much faster, so "progress per dollar" fell 63% between the first halves of 2025 and 2026. The 63% isn't a METR or Anthropic figure. It's third-party arithmetic that divides a capability index by estimated R&D spend (about $1.85 billion against $9.74 billion). METR's 5 November assessment reports 1.9x faster progress and says its benchmark isn't a cost-efficiency measure. Fact or hype: Partly fact. The METR progress figure is real; the 63% is an outside calculation built on estimates.

Why it mattersFrontier models keep improving but cost more to build, which suggests top-tier prices won't fall quickly. Budget for that rather than waiting for a price crash.

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

The big labs pile into AI for science

@kimmonismus noted that Meta is joining the push to use AI for scientific discovery, which leaves almost every frontier lab working on it. Google DeepMind's Pushmeet Kohli also hinted that upcoming releases will show AI speeding up research. Meta's research blog backs the general direction (its models helped on six maths papers in October, and it now supports university research), but I couldn't open the specific announcement the post refers to.

Why it mattersExpect more lab launches aimed at pharma, materials and engineering R&D, not just chat and coding. If you work in a research-heavy sector, watch these closely.

Sourcesx.com post ↗ · research.meta.ai ↗

AI's job impact is showing up in narrow, routine work first

Via the Wall Street Journal, @kimmonismus pointed out that US professional and business services added 70,000 jobs in October despite fears about AI and white-collar work. Document-preparation jobs, though, keep falling: US Bureau of Labor Statistics data show about 27,800 in September against 32,400 a year earlier (the post says 17%; the figures I could check point to about 14%).

Why it mattersThe pressure is on repetitive document work, not whole professions. If your team does a lot of form-filling or document prep, that's where AI will save time first, and where roles will change.

Sourcesx.com post ↗ · wsj.com ↗

reanimate: a skill that checks its own work frame by frame

@dr_cintas shared reanimate, a free, open-source Claude Code skill by Niaz Morshed. You give it an animation you like (a live site, a screen recording, an MP4, a GIF or a Lottie file). It films it, writes a timing and easing spec, rebuilds it in your stack (CSS, Framer Motion, GSAP or three.js), then films its copy and compares the two frame by frame, looping two to five rounds until they match. Install it with "npx skills add NiazMorshed2007/reanimate -g". The README reminds you that fonts, images and code still belong to their owners.

Why it mattersIt's a good example of a "verify and loop" skill, where the agent checks its own output instead of guessing once. Use it to learn how an effect is built, not to copy someone else's site.

Sourcesx.com post ↗ · github.com ↗

One agent in one chat beats a swarm, says @ctatedev

@ctatedev says that after juggling lots of agents, he's back to talking to just one agent in a single chat, and his sanity is restored. It's opinion, but a useful counterpoint to "run 20 agents at once" posts.

Why it mattersFor most small teams, one well-briefed agent beats a swarm you can't supervise. Add a second agent only when you have a clear handoff (see Agent Hub below).

Sourcesx.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.