# 📆 ThursdAI - Jan 15 - Agent Skills Deep Dive, Claude Cowork for the Masses, GPT 5.2 Codex in the API

Source: https://www.youtube.com/watch?v=qszIEZzNRsI
Recap page: https://rapidrecap.app/video/qszIEZzNRsI
Generated: 2026-01-16T02:33:23.212+00:00

---
## Quick Overview

The ThursdAI January 15 episode provided a deep dive into Anthropic's Agent Skills, celebrated the release of Claude Co-work for non-coders, confirmed the API release of GPT-5.2 Codex, and covered significant open-source model updates and industry drama, including Apple's partnership with Google Gemini for Siri.

**Key Points:**
- Anthropic's Agent Skills, previously requiring a subscription, are now an open standard adopted across multiple tools, and the show promised to detail how to use them without paying Claude.
- Claude Co-work launched, bringing Claude Code's agentic capabilities to the desktop app for non-coders, which was reportedly written entirely using Claude Code in a week and a half sprint.
- OpenAI finally released GPT-5.2 Codex via API, priced at $1.75 per million tokens, which agent harnesses like Cursor immediately used, with Cursor creating a whole browser from scratch with over three million lines of code using it.
- Byron released M3, an open-source 235 billion parameter medical LLM fine-tuned from Qwen 3, claiming it beats GPT-5.2 on healthcare benchmarks with a 65.1 score on Healthbench.
- Muan released LongCat Flash Thinking 2601, an MIT licensed model with 560B total parameters, achieving 100% of AME25 and outperforming Claude Opus 4.5 thinking on specific tasks.
- Apple announced that Google's Gemini will power Siri, running on-device and in private cloud compute, leading to speaker frustration that Siri remains 'still so so so stupid' despite the partnership.
- OpenAI announced a $10 billion partnership with Cerebras for 750 megawatts of compute, inked for 2028, highlighting the race among LLM labs to secure specialized high-speed inference hardware.

**Context:** The 'ThursdAI' episode for January 15, hosted by Alex Walov with co-hosts Misti and Wolf from Ravenwolf, focused on recent major developments in the AI ecosystem, particularly emphasizing agent capabilities and accessibility for non-developers. Key discussions centered on Anthropic's Agent Skills becoming an open standard, the launch of Claude Co-work, and major model releases from both open-source and large corporate labs, alongside notable industry drama.

## Detailed Analysis

The broadcast covered several major updates, starting with Anthropic's Agent Skills, which are now an open standard, enabling broader adoption beyond Claude Code users, with a guest promised to explain usage without a subscription. A major highlight was the release of Claude Co-work, a desktop application extending agentic capabilities to non-coders, which was developed in just a week and a half using Claude Code itself, demonstrating the power of generalized coding agents. In API news, GPT-5.2 Codex became available, leading Cursor to build a massive three-million-line browser project. In open source, M3, a 235B medical LLM, claimed superiority over GPT-5.2 on health benchmarks, and LongCat released LongCat Flash Thinking 2601, an efficient model outperforming Opus 4.5 on some benchmarks. Drama included Anthropic blocking Open Code's use of subsidized subscriptions and blocking competitor XAI entirely, alongside the surprising news that Apple chose Google's Gemini to power Siri, a move met with skepticism regarding Apple's internal AI efforts. Finally, OpenAI secured a massive 2028 compute deal with Cerebras, signaling continued investment in specialized hardware for fast inference.

### Agent Skills & Co-work Accessibility

- Agent Skills adopted as open standard
- Claude Co-work launched for non-coders on Mac subscription
- Co-work written 100% with Claude Code in 1.5 weeks
- Encouragement for non-developers to use agents for tasks like taxes and cleaning

### Open Source Model Releases

- Byron released M3 (235B medical LLM) claiming to beat GPT-5.2 on healthcare benchmarks
- LongCat released LongCat Flash Thinking 2601 (MIT licensed, 560B total params) outperforming Opus 4.5 on some benchmarks
- Black Forest Labs released Flux2 Klein (4B/9B models) for fast editing

### Big Company API and Product Updates

- GPT-5.2 Codex released via API at $1.75/million tokens
- Cursor used Codex to create a three-million-line browser from scratch
- Google shipped Personal Intelligence in Gemini, reasoning across user data with opt-in controls

### Industry Drama and Partnerships

- Anthropic blocked competitor XAI from using Claude services and blocked Open Code's subsidized subscription access
- Apple announced Gemini will power Siri, leading to criticism that Apple lacks its own competitive LLM
- OpenAI inked a $10 billion deal with Cerebras for 750MW compute starting in 2028

### Weights & Biases News

- Host announced the 'We've Hacks 3' hackathon in San Francisco on Jan 31st/Feb 1st with the theme 'self-improving agents'
- Vision and Video 3.1 updated with 4K upscaling and 9:16 vertical video support
- QTI released Pocket TTS, a tiny 100M parameter open-source TTS running 6x real-time on CPU/MacBook Air