Has Google retaken the lead? Gemini 4 Argon goes global at a cut price

Gemini 4 Argon, Google's new frontier model, rolls out globally

This is the release that resets the leaderboard for the next few weeks, and the pricing is the part to pay attention to. Gemini 4 Argon is available now through the Gemini API, AI Studio and Android Studio, in the Gemini app and AI Mode for AI Pro and Ultra subscribers, and in Google Antigravity for Ultra subscribers. Introductory API pricing is $2 per million input tokens and $10 per million output tokens, with cached input 95% off, rising to $4 and $20 after 30 November. That gives you a clear window to test it cheaply against Claude Opus 5.5 and GPT-6 Sol. Two weeks ago Google held Argon back for cyber defenders first; now everyone gets it.

Google says Argon ranks first on the Vals Index across finance, coding, legal and tax work, offers a 1 million token unified input and output limit with a new continuation protocol for long outputs, and cuts indirect prompt injection success to 0.7% on Gray Swan's benchmark. Cursor reports 52.8% on CursorBench at default effort, on a par with Opus 5.5. Every number so far comes from Google or its partners rather than independent testing. If you build agents, the long-output and prompt-injection improvements are the features most worth checking for yourself.

Liquid AI releases LFM2.5-24B-A2B, an open-weight agent model built to run on your laptop

If you want a capable agent model that runs privately on your own machine, this is one of the most interesting open releases in weeks. LFM2.5-24B-A2B is a 24 billion parameter mixture-of-experts model with only 2 billion parameters active per token, so it decodes at around 184 tokens a second on an Apple M5 Max, according to Liquid. The context window jumps to 131K tokens, the weights are on Hugging Face, and it already runs on llama.cpp, MLX, vLLM and SGLang, as well as through OpenRouter.

Liquid's own benchmarks show it beating Qwen3.6-35B-A3B on tool use and agent tasks, while trailing bigger models on harder coding and science tests such as SWE-Bench Verified. Treat it as a fast local workhorse for tool calls and routine automation rather than a replacement for a frontier model.

Facebook lets you describe what you want to see, with an LLM shaping your Feed and Reels

This is one of the first times a mainstream social feed has handed users a plain-English steering wheel. A new "Your algorithm" space in the three-dot menu lets you type the topics you want more of, and Meta says large language models interpret your intent and blend it with its usual signals, such as likes and watch time, updating recommendations within the same session. Tapping "Interested" or "Not interested" can now trigger follow-up questions so you can be more specific.

Meta says the feature is rolling out but has not given a country list, so UK users may need to wait. For anyone who has given up fighting the feed, it is worth a try once it appears.

Nvidia readies the software for RTX Spark laptops, with local agents and a new Nemotron model

RTX Spark laptops go on sale on Friday 16 October, and Nvidia used the run-up to show they will arrive with a proper local AI toolkit rather than an empty shelf. Agent support is in place for GitHub Copilot, Hermes, OpenClaw, Perplexity, Pi and WorkBuddy, Nvidia's PAIR tool is leaving beta, and a new Nemotron 3.5 Bolt 74B model is due on 15 October. There are creative updates too, including Broadcast 2.3 and ComfyUI RTX Video tools. If you have been waiting for a laptop that can run serious models and agents offline, this is the clearest sign yet that the software side is ready.

Gemini 4 Argon lands in Snowflake Cortex AI on day one

For teams whose data lives in Snowflake, this means you can try Google's new model against governed company data without moving anything out of the platform. Snowflake says Argon is available in Cortex AI in public preview, the same day as Google's launch, just as Claude Haiku 5.5 was last week. Day-one availability inside data platforms is quickly becoming the norm, which makes it easier to swap models in and out of existing pipelines as the leaderboard shifts.

Runway connects its image and video tools to Replit for web design

If you build sites with an AI coding tool, you can now generate images, video and other assets inside the project instead of bouncing between apps. Runway's MCP connection to Replit lets the Replit agent call Runway's generators directly, from first mock-up to finished site. It is a small but practical step towards coding agents that handle the whole visual side of a build, not just the code.

Apple unveils HomeView, a Siri AI smart home hub, but it will not launch in the EU

Apple's 13 October event introduced HomeView, a $299 hub with a 7-inch touchscreen for controlling the home, playing music, video calls and hands-free cooking help, launching on 13 November in eight countries. The catch for European readers is that HomeView will not be sold in the EU because Siri AI is not available there on iOS, and Apple has given no timeline for a fix. Reports also say some AI features need an iPhone or iPad nearby. UK buyers should check whether the UK is among the launch countries before planning around it.

Anthropic explains Scion, the in-house replacement for git built for agent-scale coding

This is a glimpse of what happens to developer tooling once agents write most of the code. Anthropic says merges into its monorepo rose from about 2,500 a day in May to more than 30,000 a day in September, which broke its git-based setup, so a team of seven built Scion, a server-based version control system, in five months. Checkouts now take seconds instead of minutes, and the infrastructure needs 200 times less CPU and memory.

Scion is not being released, but if your team is scaling up coding agents, the design notes are a useful early warning about where your own tooling will strain. It follows Monday's self-replication research, which went viral overnight (see Trending below).