I Swapped Claude Code's Engine for MiniMax M3: Full Power for $20 a Month

The video opens with a simple but expensive question: do you need a costly subscription to enjoy Claude Code, or can you swap the engine and get the same steering wheel for $20 a month? Erhan Meydan tries the second path and plugs MiniMax M3 under Claude Code. The roughly 28-minute video is an end-to-end experiment, from setup to a real project.

The setup follows MiniMax's official Claude Code documentation: the Anthropic base URL is pointed at MiniMax's compatible endpoint, the token-plan key is entered as the auth token, and the model name is set to MiniMax-M3. Then the claude command is run in the terminal, the folder is trusted, /status confirms the base URL points at MiniMax, and /model confirms M3 is active. Thinking mode can be toggled from /config.

The economics behind the $20 headline: MiniMax's Plus token plan costs $20 a month with roughly 1.7B tokens of M3 usage, Max is $50 for about 5.1B, and Ultra is $120 for about 9.8B. On pay-as-you-go, M3 lists at $0.30 per million input tokens and $1.20 per million output tokens at standard tier. Because agentic loops fire dozens of requests per task, the flat plan stays more predictable than metered billing under heavy daily use.

What kind of engine is M3? Released at the end of May 2026, it ships with a 1M-token context window plus tool use, caching, and vision input. Published aggregates place it in a high percentile on intelligence, strong on coding, and very high on knowledge-heavy question sets. The notable part: the version number changes capability, not price. M3 promises a more capable engine in the same price band.

But actually getting the full 1M window inside Claude Code takes one setting. A community-reported issue notes Claude Code can read model metadata through the custom backend as 200K and compact the session early. The fix is to set the auto-compact threshold to one million as part of the setup. If you plan long sessions like the ones in the video, treat that setting as mandatory, or half of the big context goes unused.

On to the tests: in the video's framing, M3 wins on planning, instruction following, and long-context work. Reading large files and making consistent edits, executing Turkish-language tasks faithfully, and turning routine coding jobs at low cost are its strong side. Independent scores back this picture: high coding and knowledge-test rankings suggest M3 stands close to frontier models on everyday developer workloads.

The areas where it trails are discussed honestly too. Fine aesthetics, pixel-level design judgment, complex animation, and risky multi-file refactors still belong to the most expensive frontier models. That honest balance matters: M3 is not the best at everything, but the number of jobs where it is the most cost-effective is large. The claim is price-performance leadership, not absolute supremacy.

The explainer in the middle clarifies three layers of the Claude Code world: MCP provides the connection, skills teach the method, and plugins ship both as a product with commands and subagents. MCP brings live data, the skill holds the recipe, and the plugin reduces installation to one step. Teams that set up this trio gain a model-independent working discipline.

The closing real project makes it concrete: an app collecting regional dialect words on a map. Data entry, listing, filtering, and a map view form an end-to-end product skeleton built inside the M3-powered Claude Code session. Real-world details like Turkish characters, local dialect data, and map interaction give it the feel of a living project rather than a synthetic demo.

The bottom line is clear: Claude Code plus MiniMax M3 with the $20 plan tears down the subscription wall and opens near-frontier coding to a wide audience. My lesson: invest in the harness and workflow first, because renting the engine is now the cheapest part. Keep frontier models in reserve for the most critical pixel work, and run everything else on this combination.

minimax m3 · claude code · token plan · agentic coding · mcp