I Cancelled Claude Code. Here's What I Use Now

Summary

The presenter explains why they cancelled Claude Code, despite considering it the best coding agent they had used. Their reasons include usage limits and cost, as well as concerns about depending on a service whose terms and availability can change. They say Claude refused to help with research on a proposed AI-model architecture, and cite Anthropic’s terms and safety documentation as examples of restrictions that may affect legitimate AI or cybersecurity work.

Claude Code: Cost, Limits, and Control

• The presenter says Claude Code’s paid plans start at $20 per month, with more expensive tiers, and that usage caps make it hard to predict when they can work.
• They argue that limits can lead developers to focus on managing tokens rather than coding.
• They emphasize that users may own code produced with the tool, but access to the service remains subject to the provider’s terms.
• The video also discusses OpenAI’s announcement about a Navier–Stokes result and a mathematician’s account of an authorship offer. The presenter uses the episode to raise questions about credit and ownership when research is done with proprietary tools, while noting that the proof had not been independently verified.

Open Models and Dropstone

The presenter considers open models such as DeepSeek, Kimi, and GLM, arguing that they may be “close enough” for many coding tasks at a lower cost. They say that running large models locally can require substantial memory and expensive hardware, and that a practical coding agent also needs project memory, integrations, security controls, and ongoing model updates.

Their chosen alternative is Dropstone, a terminal-based coding agent that uses open models. The presenter describes it as offering different performance modes, with the ability to swap models as better open options become available. In their setup, the models run on US-hosted servers rather than being downloaded locally.

Security Discussion and Demonstration

The video argues that agent safety should not rely on a model policing itself: risks can come from commands an agent executes, regardless of where the model comes from. The presenter describes five Dropstone safeguards:

1. Commands are checked before execution, and dangerous commands are blocked even in “accept all” mode.
2. On Mac and Linux, commands can be sandboxed to restrict writes to the project.
3. Web content is checked for hidden instructions aimed at the agent.
4. Risky actions require a request in the user’s own words.
5. The agent will not send data to an address found only in a web page.

To test these claims, the presenter uses a disposable Linux machine with no access to personal files. In the demonstrations, Dropstone declines to delete planted files, blocks a curl | bash installer command, and identifies a hidden prompt-injection attempt to exfiltrate a planted SSH key.

Head-to-Head Coding Test

The presenter gives Claude Code and Dropstone the same four tasks in fresh copies of a project: implement a full-stack feature, fix three bugs, build a pricing page from a specification, and test whether each agent remembers a project rule across sessions. Hidden tests are used to assess the coding tasks.

• Results: Both agents pass all 27 hidden tests.
• Reported API cost: Claude Code costs 20–38 cents per task; Dropstone costs under one cent to four cents. The presenter says the full-stack feature was about 25 times cheaper with Dropstone.
• Speed: Claude Code completes tasks in about a minute, while Dropstone takes one to three minutes. Its pricing-page task takes almost 13 minutes, partly because it opens a browser and checks its work.
• Memory: Both remember the user’s rule in a new session. The presenter also notes that Dropstone saved and reused a lesson about the project’s tests without being asked.

The presenter says they used Dropstone’s cheapest “fast” mode for all four tasks and did not need its heavier mode. They conclude that smaller or cheaper models may be sufficient for many routine tasks, though Claude Code remains the faster choice in this comparison.

Overall Conclusion

The presenter recommends considering open-model tools for developers who frequently hit Claude Code’s limits, want lower costs, or want more control over their tooling. They still describe Claude Code and Opus as the strongest option for users who can afford it and do not encounter limits. The central argument is to avoid building an entire workflow around a tool whose access and terms the user does not control.

Reviews, Guides, and Tests Covered

• Product comparison: Claude Code versus Dropstone on four coding tasks, with hidden tests and reported costs and completion times.
• Security demonstration: A live attempt to trigger unwanted deletion, run an unreviewed installer, and exploit a prompt injection.
• Practical guidance: An explanation of the hardware and software components involved in running open models locally, along with safeguards to look for in coding agents.

Main Speaker and Sources

• Main speaker: The Epic Programmer presenter, speaking from personal experience and describing their own tests.
• Sources mentioned in the video: Anthropic’s consumer terms and developer/sandboxing documentation; NIST’s evaluation of DeepSeek agents; research on sleeper agents and prompt-injection defenses; reporting on OpenAI’s Navier–Stokes announcement and controversy; and examples involving Replit and Google’s Antigravity.