Streaming speaker extraction shrinks: Falcon uses 0.5M parameters
A new streaming audio-visual model uses lip video to isolate a voice, and the paper reports 11 times lower GPU latency than a larger rival.
770 stories
A new streaming audio-visual model uses lip video to isolate a voice, and the paper reports 11 times lower GPU latency than a larger rival.
A skill is a folder with one SKILL.md file. Here is where to save it, what goes in the frontmatter, how to test it and what to check when Claude ignores it.

Google lists $2 input and $10 output per million tokens as an introductory price, then $4 and $20. It has not said when that changes, and the pricing page does not list the model.

Mistral's open-weight Le Chonk costs a fraction of Claude Opus 5.5 per token, but Artificial Analysis scores it 32 points lower on its Intelligence Index.

Amazon's shopping assistant can now skip Subscribe & Save deliveries, check refunds and edit lists by chat, for all U.S. customers.
The open-weight model, nicknamed Le Chonk, costs $0.70 per million input tokens and is tied with GLM-5.3 on a coding agent index, Mistral says.

Fill with Gemini shows the pages behind every generated value and handles up to 3,000 cells per request, Google says.
In its Universe post, GitHub also says it is building a new Git architecture and moving to a multi-region Azure deployment by early next year.
Cursor starts at $8 a month as a standalone editor, Codex comes with ChatGPT plans from $20. Here is what each includes, from their docs.
Open Billing and click AI usage to see credits used, extra usage and a per-model breakdown, as of October 29, 2026.

Alibaba released Qwen4 Max on November 16, 2026 after a five-day preview. Smaller open-weight Qwen 4 models are promised, with no names, sizes or date.
Run claude --continue for the last session or claude --resume to pick one. Plus naming, forking, what gets restored and fixes for missing sessions.
A Peking University and JD.com paper adds a 57M-parameter module to a 4B vision-language model and reports top scores on four simulated robot benchmarks.
Carbon edges Opus 5.5 on Artificial Analysis, yet it is limited to some cyber defenders while Opus 5.5 costs $4 and $20 per million tokens for everyone.
Amazon holds its Q3 call at 2:00 p.m. PT today, with AWS growth and the guidance range of $197B to $202B in sales in focus.
Ads in AI Mode are scaling in India, Canada, Australia and seven other markets, with Brazil, Mexico and the UK planned for Q4 2026.
ServiceNow beat its own guidance, but a fourth-quarter outlook below analyst estimates sent shares down about 18% after hours.
Run the same command you used to install Codex. Homebrew uses brew upgrade, and the desktop app updates itself on macOS and Windows.

Both come with a paid Claude plan. Claude Code is for software work in a repository, Cowork for documents, files and research, with no terminal.

Google began a staged rollout on October 28 to a subset of cyber defenders. Developers and consumers get it "soon", with an introductory price of $2 per million input tokens.
Switch models mid-session, at launch, with an environment variable or in settings, and find out which one wins when they disagree.
A new benchmark gives nine agent systems one hour to write programs that render a finished video. The best met every core requirement on 76.3% of tasks.
The Korean inference chip maker will scale its RNGD chip, already in mass production, and fund a third-generation design with Broadcom.
Claude's SDK runs Claude models in your own process. Cursor's runs on its hosted models and can deploy agents to its cloud. Here is how they differ.