🤖 AI Weekly Recap (Week 7)

Happy Sunday! We just had another crazy week in AI. China just dropped the world’s most advanced open source AI model while there is a new AI that lets you create entire agentic node workflows from text instructions.

And that's not all, here are the most important AI moves you need to know this week.

1. ByteDance unleashes Seedance 2.0

TikTok parent ByteDance unveiled Seedance 2.0, a new multimodal AI video generator designed to create “cinematic content” using text, images, video, and audio, now rolling out to select users on its Jimeng AI platform.

• Generates 2K video with ~30% faster output than Seedance 1.5

• Supports seamless video extension, precise references, and natural language control

• Reportedly beats Sora 2 and Veo 3.1 in practical testing (per CTOL)

• Outputs are completely watermark-free, unlike rivals

Try it Now → https://seed.bytedance.com/en/seedance2_0

2. Anthropic Releases Cowork on Windows

Cowork, Anthropic’s agentic workspace built on Claude Code, is now available on Windows with full feature parity. It’s designed to let Claude act more like a real coworker, planning, executing, and managing multi-step work across your files.

• Full Windows support with file access, plugins, MCP connectors, and multi-step tasks

• Global & folder instructions let you set tone, format, and work style once and reuse it

• Claude can read, edit, create, and organize files with higher autonomy than chat

• Available in research preview for Pro, Team, and Enterprise plans

Try it Now → https://claude.ai/

3. Gemini 3 Deep Think Breaks Reasoning Benchmarks

Google’s specialized reasoning model, Gemini 3 Deep Think, scored 84.6% on ARC-AGI-2, beating GPT-5.2 and Claude by a wide gap. ARC-AGI-2 is designed to resist memorization and test real, on-the-fly reasoning.

• Gemini 3 Deep Think: 84.6%, GPT-5.2: 52.9%, Claude Opus 4.6: 68.8% (humans ~60%)

• Also hit 48.4% on Humanity’s Last Exam and 3,455 Elo on Codeforces (elite human tier)

• Achieved gold-medal-level results in 2025 math, physics, and chemistry Olympiads

• Verified inference cost is $13.62 per task, raising scale and economics questions

Try it Now → https://gemini.google.com/

4. MiniMax Releases M2.5 Coding Model

MiniMax unveiled M2.5, a 100% open-source model with just 10B active parameters that rivals closed giants like Opus and GPT-5.x, at a fraction of the cost.

• Competitive with Opus 4.6 / GPT-5.2 in coding, agents, search, and office work

• Runs at up to 100 tokens/sec, ~3× faster thinking efficiency than Opus

• $0.3/M input, $1.2/M output, up to 20× cheaper than frontier models

• SOTA results on SWE-Bench, BrowseComp, and real-world agentic tasks

Try it Now → https://agent.minimax.io/

5. Krea AI Launches Node-Based Automation

Krea AI now lets creators visually chain multiple AI tools into a single workflow. Instead of jumping between apps, you can design, automate, and run entire creative pipelines inside one interface.

• Node-based canvas to connect image, video, and AI tools step by step

• Run full workflows with a single click, no platform hopping

• Easily tweak or swap tools inside a pipeline without breaking it

• Built for creators who want speed, consistency, and scale

Try it Now → https://www.krea.ai/features/nodes

6. Z.ai Releases GLM-5 With Record-Low Hallucinations

GLM-5 is the latest open model in z.ai’s GLM series, released under an MIT License and built for real-world, agentic knowledge work. It now leads the industry in knowledge reliability, outperforming Google, OpenAI, and Anthropic on hallucination benchmarks.

• Achieves a -1 score on the AA-Omniscience Index, the lowest hallucination rate ever recorded

• Uses a massive 744B-parameter MoE architecture with only 40B active per token for efficiency

• Introduces “slime,” a new async RL system that speeds up agentic training and long-horizon reasoning

• Priced at ~$0.80 input / ~$2.56 output per 1M tokens, ~6–10× cheaper than Claude Opus 4.6

Try it Now → https://chat.z.ai/