Hello everyone, welcome to the latest AI news for November 10th. On screen are the highlights of this episode. Logan is hyping Patrick again, shouting about GEMINI — GEMINI 4 Carbon is about to become officially available. Based on past episodes, every time he shouts about GEMINI, a new usable model launches 13 to 17 hours later, meaning it's coming early tomorrow morning. Carbon has been in an unavailable state for over ten days, and right after being knocked off the top by Fable 5.2, it's about to become available again — it's really awkward. Of course, Carbon isn't directly competing with Fable, because it's much cheaper than Fable.

Latest news: GEMINI 4 Carbon is already live on Vertex AI. The EAP (Early Access Program) for Gemini 4 Hydrogen has officially started, with a release expected soon. Given Google's usual style, it will probably wait until next week to launch. According to Arena testing, Hydrogen is basically on par with AGON but much faster, possibly reaching 300–500 TPS. Most importantly, for low-price PR users, this Flash model may be usable in Antigravity.

Latest news: Gemini Hydrogen seems to have started gray testing. The suffix of the 3.8 Flash model in Antigravity has changed from N to S, and the TPS has increased a lot, averaging nearly 200 TPS. Check whether you've been routed to it.

Meta has released Muse Amber, the smallest and most efficient model in the Muse series, focused on complex reasoning execution and everyday agent work. From the benchmark charts, it targets Haiku 5.5 and 6. Except for losing to Haiku in software engineering, LUNA leads in almost all other benchmarks. On the Artificial Analysis Intelligence Index, it scores 36, on par with Haiku 5.5, with a cost per task of only $0.14 — one thirteenth of Haiku 5.5's — making it currently the cheapest model. With 36 points of intelligence plus the lowest price, did Meta just farm benchmarks again, or is there real substance? We'll see from actual usage experience. Muse Amber is now live on the Meta Model API, priced at $0.10 per million input tokens and $0.50 per million output tokens.

Musk has finally responded on the progress of GROK 4.8. He said pretraining is complete and it's now in the reinforcement learning stage. Later this week, it will first be opened for testing to some GROGBT users. Keep an eye on whether the bot suddenly gets better. This time he was finally honest and didn't hype it up, claiming 4.8 is invincible — he said 4.8 doesn't count as a leap, but the foundation is very solid.

On Reddit, the Kun community has a post saying Kun 4 will be released on November 11th at 3 PM. Honestly, I almost forgot Double 11 is almost over. An Alibaba Double 11 release of an Alibaba large model does seem possible. Although this is most likely a junk post, you never know.

According to disclosures in the MATHS 5.2 system card, it is the same model as "model 3 early" tested in M1T2. From M1T2's previously disclosed investigation, there are also model 3 and model 4 that are stronger than early. But MET2's report was actually based on an investigation from September 23rd. What ANTHROPIC released may be model 3, so Fable 5.5 might be model 4. After all, the checkpoints previously speculated to be Fable 5.5 produced very impressive outputs, clearly and noticeably much better than Fable 5.2. ANTHROPIC is probably selectively releasing model capabilities.

ANTHROPIC partnered with Dio Computers to have Cloud Fable 5.2 design three kinds of devices from simple prompts — a hands-free speaker, a voice assistant hub, and an electronic pet — including PCB layout, CAD modeling, firmware writing, and DFM checks. The three circuit boards were manufactured as prototypes directly by a factory from the design files, and they still worked normally after weeks of use.

Dot, besides having its own cloud mini-computer, can now connect multiple local computers to the same Dot and choose which computer executes each task — for example, having Dot build and test apps on a dedicated computer while you keep coding on your own laptop.

OpenAI launched the Codex game plugin and a custom handheld console. Through the ModRetro Chromatic plugin in Codex, you can design levels, create sprite animations, and automatically build ROM game image files, then try them on the handheld. The Codex special edition Chromatic handheld is now available on Supply.co, but unfortunately it doesn't ship to China — previous Cloud merchandise could be delivered there. Below is a detailed showcase of the handheld; take a look if you're interested.

Graph's research based on Similarweb data shows ChatGPT accounts for over 81% of total traffic among major global AI chatbots, with no rival reaching 10%. OpenAI developed a system called Control Room to track model tasks encountering safety issues during internal research. When researchers or monitoring systems detect problems, they can use it to pause or stop the relevant tasks. OpenAI says that from the moment an intervention decision is made, the system can stop the task within 35 seconds and continue obtaining model outputs. The work it covers includes model training, evaluation, and data processing.

After GEMINI 4 CRYPTON appeared in Design Arena, ZENON and RADON have also appeared on Design Arena. Does Google just love naming things after noble gas elements? Krypton, Xenon, and Radon all belong to the same group of noble gas elements, but Carbon isn't, and neither is Hydrogen — hard to say what Flash really means.

Google announced an expansion of its eighth-generation TPU product line, adding two new systems: TPU 8TX and 8 Max. Together with the earlier 8T and 8I, four eighth-generation TPU systems have been released within one year. The 8TX is optimized for secure and confidential inference, designed specifically for sensitive enterprise workloads — dedicated to sensitive data in banking, government, and similar sectors, with dual chips.

The 8i Max system doubles compute, memory, and bandwidth, and is purpose-built for million-token agents. Google Antigravity has launched a brand-new knowledge work interface layout, optimized for research, writing, and producing deliverable content. It retains the same agent framework, plugins, and MCP tools, while removing interface elements such as the terminal and Git. You can select the interface layout in the General tab of the settings.

In AI Studio, users can now build AI applications powered by various Google Maps features—including 3D buildings, Street View, and weather—using just simple prompts.

Thinking Machines' Inkling ARS small model is available for free on OpenRouter. LongCat 2.5 is now free on OpenCode for two weeks, with zero data retention. MiniMax M3.1 Flash is now free in OpenCode. DPAI announced that DeepSeek V4.1 Flash 1017 is now available for free through DPAI Code. Let me take a look at what this is about—it claims to be 100% free and ad-supported.

A total of 18 models are available, as shown in the image. The nine on the left are free with limits: 200 per day on the free tier, and 400 on the paid Pro tier. The nine on the right are completely free. Has anyone tried them? Feel free to share how the speed is and whether they are quantized.

Meituan's LongCat has officially been open-sourced. LongCat 2.5 uses the MIT license and is a natively multimodal model designed for long-horizon agent tasks, with 1.6T total parameters and approximately 48B activated, supporting a 1M-token context window.

That's it for this episode of AI news. See you next time—follow me so you don't get lost!