The Day in AI
In short
- Google DeepMind released Gemini 4 Neon on 18 November, a mid-tier model that scores 61 on the Artificial Analysis Intelligence Index.
- Google added computer use to the Gemini app on macOS, so it can work across local files, desktop apps and the web in a background window.
- Google put Tool Retrieval in preview in the Gemini Interactions API, so agents load a tool's full schema only when they need it.
- RadixArk released its first model, the open-weight DaisyRose-27B, and reports it doubles its base model's Terminal-Bench 4.1 pass@1 score.
- Anthropic raised $81.3 billion in its IPO, according to the South China Morning Post and The Straits Times.
- Meta plans to sell its pocket-sized Muse Charm device for $300, according to The Information.
The day in brief
Google DeepMind released Gemini 4 Neon on 18 November, a mid-tier Gemini 4 model that scores 61 on the Artificial Analysis Intelligence Index at $3 per million input tokens.
Gemini 4 Neon sits between Gemini 4 Carbon above it and Gemini 4 Hydrogen below it. Artificial Analysis measured 61 on its Intelligence Index at the top thinking level, called xhigh, ahead of Claude Opus 5.5 at 59 and GPT-6.2 Sol at 57. Google says that puts Neon third overall. On the Vals Index, Google's own comparison table gives Neon 70.9% against 70.1% for Opus 5.5.
Neon is priced at $3 per million input tokens and $15 per million output tokens. It has a 1 million-token context window and four thinking levels, which set how long the model reasons before it answers, so a higher level spends more tokens on harder problems. Google reports that at the default high setting, a task on Neon costs 32% below the same task on Carbon. In Google Antigravity, Neon can run several Hydrogen models in parallel as subagents and hand each one part of a job.
Google also shipped two agent features on 19 November. The Gemini desktop app on macOS can now fill in forms, sort files into folders and build slideshows. It works in its own background window with a live preview in the chat, and users choose which apps it can reach and can pause, stop or approve actions. It needs a Google AI Plus, Pro or Ultra subscription. In the Gemini API, Tool Retrieval lets developers mark tools as deferred. Google says the model still sees each deferred tool's name and description and loads the full parameter schema only when it asks for that tool, all within one request, so large tool libraries stop filling every prompt.
RadixArk released DaisyRose-27B on 18 November, a fine-tune of Qwen3.8-27B trained with reinforcement learning (RL) on about 1,200 long-horizon agent tasks. RadixArk reports Terminal-Bench 4.1 pass@1 rose from 5.6% to 12.0%, averaged over eight runs in its own harness. Training took 18 days on 144 H200 GPUs, and RadixArk says its internal backend reached 37% model FLOPs utilization (MFU) against 16% for the open-source trainer. Weights are on Hugging Face in BF16, FP8 and NVFP4.
Also in the news
- Anthropic raised $81.3 billion in its IPO, which the South China Morning Post and The Straits Times report is larger than SpaceX's record debut.
- Meta plans to sell Muse Charm for $300 with up to 12 months of free Verizon 5G data, according to The Information.
- Superhuman made its Go assistant the central agent in Docs and Mail, added 20 third-party agents, and says Go has over 1 million active users.
- OpenAI added Live Dictation to ChatGPT, which shows the transcript as you speak, and removed the 10-minute limit on voice dictation.
- Google added a sidebar to the Search homepage on desktop that lists and continues AI Mode chats.
- AMD closed its $8.2 billion acquisition of World Labs, according to Yahoo Finance.
- Sierra, Bret Taylor's customer service startup, is in talks to raise $1.5 billion at a $30 billion valuation, The Information reports.
- Alibaba is reportedly training its next Qwen model on its own chips.
- Boltz, a London drug discovery startup, reportedly raised a $110 million Series A led by Bessemer.