Heidelberg AICurriculum

Claude Code Is Too Expensive. Use This Instead

Leon van Zyl ·2026-05-28 ·1 min read

Summary written by us from the video's transcript. The video, and everything in it, is Leon van Zyl's work.

Learn how to replace Claude’s coding models with the open‑weight Minimax M2.7 in Claude Code, set up authentication, and run a full Reddit‑clone project at a fraction of the cost.

Takeaways

  • Minimax M2.7 can replace Claude’s coding models with comparable quality at roughly one‑tenth the price.
  • Configuration requires only a settings.json change and an API key; no workflow changes are needed.
  • Requests are counted per call, not by token usage, giving predictable cost control across simple or large prompts.
  • Claude Code skills (design, browser, auth) can be installed and used with Minimax to automate full‑stack development.
  • Sub‑agents enable parallel feature implementation, accelerating project completion while staying within a modest budget.

Why Switch to Minimax M2.7

Leon demonstrates that Minimax M2.7 delivers performance comparable to Claude Opus (Sweetbench Pro 4.6) while being roughly ten times cheaper, making it attractive for developers who find Claude’s pricing prohibitive.

Configuring Claude Code to Use Minimax

Create a **settings.json** in your Claude project folder and point the `anthropic.baseUrl` to Minimax’s endpoint, inserting your Minimax API key. The repository with the exact JSON snippet is linked in the video description.

Getting an API Key and Choosing a Plan

Sign up at minimax.io → API Platform → Token Plan page. Most users will select the **Starter**, **Plus**, or **Max** plans; the Plus plan provides 4,500 requests per five‑hour window, independent of token count. Requests are counted per call (e.g., a simple hello or a full PRD each counts as one request).

Running the Reddit Clone with Claude Code

Start Claude Code in *planning mode*, paste the provided tech‑stack prompt, and let Minimax generate a project plan in under two minutes. Switch to *change mode* to see token usage update instantly. Use the **goal** command to keep the agent looping until all features—authentication, UI design, testing—are complete.

Leveraging Agent Skills and Sub‑agents

Install skills such as Front‑End Design, Agent Browser, Better Auth, and Next Best Practices via Claude Code’s skill manager. Sub‑agents run in parallel to implement different features, while the browser skill performs end‑to‑end visual testing of the UI.

Cost Efficiency and Usage Monitoring

With a $20 plan, Leon used only 11% of his quota despite running sub‑agents, skills, and automated tests—showing that Minimax’s request‑based pricing can handle complex workflows affordably. Usage details are visible on the token‑plan page under “current text generation usage.”

How Minimax Trains Its Models

Minimax M2.7 was trained using its predecessor to help construct the next version, a process they describe as the model participating in its own evolution. Independent benchmarks (MLE bench) placed M2.7 second only to Opus and GPT‑5.4.