Summary of Key Points
OpenCode, a phenomenon-level open-source project on GitHub with 160,000 stars and 7.5 million monthly active developers, recently released version 2.0, which represents a complete overhaul: the API has been restructured, the technology stack has been switched from Bun to Node.js to address memory issues, and the desktop version has been migrated from Tauri to Electron. New features include multi-session parallelism and a cross-device agent network. The team has improved product details through an “excessive design” approach using tokens and shared their insights into the AI industry (model routing is overestimated, inference profit margins exceed 90%, and local models do not necessarily save costs), as well as the trend towards voice-based, hands-free AI interactions.
OpenCode 2.0: Three Major Core Changes
Why a Complete Rewrite? Founder Dax believes that “everything needs to be done three times correctly”: Version 0 was a prototype, Version 1 was for validating requirements, and Version 2 is a complete redesign based on expert knowledge.
- API Reengineering: The previous API was developed organically (adding whatever came to mind), but now it’s a well-designed, unified system to avoid confusion.
- Technology Stack Upgrade: Moving from Bun to Node.js solved the issue of high memory usage (over 2GB). Bun had dedicated APIs that made servers dependent on it; removing these APIs reduced memory consumption significantly.
- Desktop Version Migration: Version 1 used Tauri, which was lightweight but slow in rendering. Now, Electron is used with Chromium for better performance and consistency, and the server can run directly within the built-in Node process of Electron without the need for an additional CLI.
User Experience: No More Waiting in Lines, Easier Cross-Device Control
The Biggest Pain Point Solved: Waiting in lines for AI tasks to complete. Now, multiple tabs can be opened, with each tab corresponding to an independent AI session. For example, one model can generate an HTML website while another creates a member list—both working simultaneously without interference, and different models can be compared.
Cross-Device Agent Network: With version 2.0, OpenCode runs as a service in the background by default, allowing all devices (computers, phones, remote servers) to connect. For instance, you can use OpenCode on a remote Linux server to control a local Mac to send an iMessage. You can even include all devices in an “agent network” and instruct it to, say, “buy a gift for my wife” (although Dax’s wife almost divorced him over this feature, but the functionality is real).
Development Methodology: Spending Money on “Excessive Design”; Underlying Tools Are Key
Why a 5-Fold Increase in Token Usage? Not waste, but “luxury design.” Previously, creating an API for file reading might have involved choosing the best of 1-2 options. Now, with AI models, all possible solutions, precedents from other products, and response mechanisms are considered—like having countless consultants help optimize the design. Although it costs more in tokens, the product details are much better.
Investing in Fundamental “Infrastructure Tools”: For example, they developed their own OpenTUI framework (written in Zig) to ensure that terminal applications (TUIs) work smoothly on all platforms. This is like laying a solid foundation before building a house; even with AI assistance, good underlying tools are essential for creating useful products.
“The Truth” About the AI Industry
- Model Routing: Overrated. Dax believes it’s better to have a “master model” that directs subordinate models. Switching models mid-session resets caches, which is costly. A “command mode” using a more expensive master model to assign tasks to cheaper subordinate models (e.g., for code modification) is both intelligent and cost-effective.
- Inference Profit Margins Exceed 90%: OpenAI/Anthropic’s inference profit margins are around 90%—earning $100 with only $10 in cost, allowing prices to be reduced by tenfold. Local models don’t necessarily save costs; cloud efficiency is 10 times higher, so the cost per token decreases significantly in the cloud. The value of local models lies in privacy (data remains on the device).
- Why Does Anthropic “Ban” OpenCode from Using Claude? Anthropic is business-oriented and wants users to transition from consumer products like Claude to paid enterprise solutions. If users use Claude through OpenCode, they might switch to other models, breaking this conversion chain. OpenAI, being consumer-oriented, is willing to spend money to encourage more usage.
The Future of AI Interactions: Voice Input Becoming the Standard
- The Team Doesn’t Even Use Keyboards: Dax’s team uses voice for Discord conversations because LLMs can understand messy inputs (such as slips of the tongue or nonsense), resulting in higher-quality outputs. Some even use footpads to trigger voice input, freeing up their hands.
- Why Voice Is Popular? It’s more natural to just speak your commands—e.g., you can ask AI to record your workout and it will automatically store the data in the database, reminding you to adjust your technique next time.
Conclusion
OpenCode 2.0 has significantly improved performance and user experience by making major changes. The team has transformed their development approach using AI and revealed the “unspoken rules” of the AI industry: high profit margins in inference, misconceptions about model routing, and the true value of local models. Future AI interactions will be more natural (with voice as the primary method), and cross-device agents will become the norm. For ordinary users, Version 2.0 eliminates the need for high memory usage and allows simultaneous execution of multiple AI tasks, greatly enhancing the overall experience.