👾 AI Nuggets #18: From $20K Agents to Free Alternatives

Welcome to edition #18 of the AI Nuggets. In this issue:

• Manus AI agent from China

• OpenAI plans to sell $20K AI agents

• Mistral OCR technology

• Alibaba’s QwQ-32B open-source model

• Build open-souce AI agents

Latest News

🔬 Tech Spotlight: Manus AI Agent from China

China has unveiled Manus, a breakthrough autonomous AI agent that promises to execute complex tasks from start to finish with minimal human input. Unlike conventional chatbots that merely generate suggestions, Manus plans and completes real-world assignments independently.

What's New

Manus stands apart from existing AI tools through its complete autonomy. It can browse websites, use multiple tools, and display its workflow in real-time. The system handles diverse tasks like creating detailed travel itineraries, analyzing stocks, building websites, and comparing insurance policies. In a demonstration, presenter Ji Yichao emphasized that Manus "bridges the gap between conception and execution." The app is accessible on an invitation-only basis but will likely become freely available soon. According to its developers, Manus outperforms OpenAI's Deep Research on the GAIA benchmark, a third-party measure for general AI assistants.

Business Impact

The emergence of Manus could reshape how businesses approach routine and complex tasks. It can reduce overhead by automating processes that currently require human attention. Think of resume screening, financial analysis, or supplier sourcing - all handled without constant supervision. The timing is notable, coming shortly after DeepSeek's R1 reasoning model and Alibaba's latest AI offerings, highlighting China's accelerating innovation in artificial intelligence.

😮 OpenAI Plans to Charge Up to $20,000 a Month for AI Agents

OpenAI is set to launch specialized AI agents designed to automate complex workplace tasks, with price points ranging from $2,000 to $20,000 per month based on capabilities.

What's New

A "high-income knowledge worker" agent will be available at $2,000 monthly, while a software developer agent will cost around $10,000 monthly. The most advanced offering, priced at $20,000 monthly, will support PhD-level research tasks. These agents differ from current AI tools by working autonomously within company systems. They can access databases, understand company operations, and execute complex workflows independently. OpenAI has already secured substantial interest, with investor SoftBank committing $3 billion to these agent products this year alone.

Business Impact

As SoftBank CEO Masayoshi Son put it, the difference between businesses with and without Enterprise AI Agents will be like "comparing a country with electricity to one without electricity." Tasks like financial reporting, document creation, and customer service could be handled more efficiently with less human intervention. SoftBank itself plans to use Cristal intelligence to automate over 100 million internal workflows. However, the substantial price points suggest these tools will initially be accessible primarily to larger enterprises, potentially creating new competitive divides based on AI adoption capabilities.

📄 Mistral OCR: World’s Best Document Understanding

Mistral AI has launched Mistral OCR, a groundbreaking document understanding API that sets new standards for accuracy and comprehension. This technology excels at processing complex documents with images, tables, equations, and various layouts, offering superior performance at 1000 pages per dollar.

What's New

It transforms images and PDFs into ordered, interleaved text and images with unprecedented precision. The system outperforms competitors across all benchmarks, particularly in handling mathematics, multilingual content, scanned documents, and tables. The technology processes up to 2000 pages per minute on a single node, making it the fastest in its category. It supports thousands of scripts, fonts, and languages worldwide, ensuring true global accessibility. Mistral OCR also introduces document-as-prompt capability, allowing users to extract specific information and format it in structured outputs like JSON for seamless integration with other systems.

Business Impact

At 1000 pages per dollar (or double with batch processing), companies can digitize entire document libraries without breaking the bank. The technology eliminates manual data entry from invoices, contracts, and forms, freeing staff for higher-value tasks. For businesses serving diverse communities, the native multilingual support opens new markets without additional investment in translation services. Technical teams can integrate document content directly into their workflows through the structured output options. Organizations handling sensitive information can self-host Mistral.

🇨🇳 Alibaba’s QwQ-32B AI Model: Small & Fast

Alibaba has unveiled QwQ-32B, a compact AI reasoning model that matches the performance of much larger models while requiring less computing power. The new model excels at math, coding, and general reasoning tasks. It's available as open-source software under an Apache 2.0 license, making it free for both commercial and research use.

What's New

Despite having only 32 billion parameters compared to DeepSeek R1's 671 billion, it delivers comparable performance. The model uses reinforcement learning to improve its reasoning abilities. It can think through problems step by step, question its own answers, and refine its responses. This approach has helped QwQ-32B perform well on challenging benchmarks for math reasoning, coding, and instruction following. The model works with a 32,000-token context length.

Business Impact

QwQ-32B can run on consumer-grade hardware, dramatically reducing deployment costs. This opens doors for companies looking to implement AI solutions for complex problem-solving, coding assistance, data analysis, or customer service automation. This means businesses can customize the model for their specific industry needs without proprietary restrictions. For technical teams, this represents an opportunity to build specialized applications with advanced reasoning capabilities at a fraction of the cost of larger models.

AI Use Cases and Tools

🪿 Automate Workflows with this Open-Source AI Agent

Engineers and developers are constantly seeking ways to streamline repetitive tasks. Enter Goose, a powerful open-source AI agent that's changing how we approach automation. This local-first tool allows you to automate complex engineering workflows with remarkable flexibility and control.

What Makes Goose Special

It works with virtually any LLM, including open-source models via Ollama integration. This means you can harness models like Alibaba's QwQ-32B without being locked into proprietary systems. The agent runs locally on your machine, keeping you in control of both execution and your sensitive data.

What truly sets Goose apart is its seamless integration with Model Context Protocol (MCP) servers. This capability allows Goose to connect with any software that adheres to the openAPI standard. The result? You can create sophisticated automated workflows that span multiple applications and services.

Real-World Applications

The applications range from navigating a computer to setting up dev environments automatically. It can scrape the web, integrate with GitHub, Figma, Slack, and other productivity tools.

Coming Soon: Goose Tutorial

MadebyAgents is currently working on a video tutorial that demonstrates how to leverage Goose with MCP servers to create powerful automated workflows. This step-by-step guide will show you how to configure Goose, connect it to various services through MCP, and build end-to-end automation solutions that save hours of manual work. Stay tuned!

That’s It for This Week

✨ Before You Go:

We’d love to hear what you think. Please share your opinion.

See you next time!

Tobias from MadebyAgents