MiniMax M3 Review: Cheap Daily Driver for AI Agent Loops — Superbash Learn

Superbash Learn

Click to load the YouTube player

Key takeaways

Use MiniMax M3 only as the execution layer after Opus 4.8, GPT 5.5, or Z has produced the plan — the creator calls those three "smarter" for "large functions and planning out."
Run agent loops and overnight build iterations on the $20/month MiniMax tier; the creator says it is already "good enough already" for agent work without paying for the $50 plan.
Point Claude Code at the Anthropic-compatible MiniMax base URL to keep MCPs and agent teams working — this combo accounts for the bulk of the creator's token burn.
Skip GLM for high-volume loops: the creator burned 20% of his weekly MiniMax quota in a single hour on GLM, and described its pricing as having "went up quite aggressively."
Treat the 5-hour rolling limit, not the weekly cap, as the real ceiling — the creator has to "really push" to hit it, but automated hourly loops can chew through it overnight.
Do not rely on M3's in-app token counter for cost tracking; the creator notes "they don't tell you exactly how many millions of tokens you used" on the hourly view.
Batch the 3 included videos into a single "column job" rather than generating ad-hoc — the creator calls the per-video cost "expensive" and plans to script the full set at once.

Video summary

Companion notes

MiniMax M3 is the cheapest daily-driver execution model in the creator's stack, but only after a smarter model has planned the work.

Pricing and Quota

The creator runs the $50/month MiniMax plan and has been "using it very very aggressively" — burning "hundreds of millions of tokens per day" without hitting the weekly cap. The bottleneck is a 5-hour rolling limit, and the creator admits they have to "really push myself to really burn that many tokens." For agent loops and development, the $20/month tier is already "good enough already."

Error Rate vs. M2.7

M2.7 "made a lot of errors" and required constant babysitting. M3 in the creator's two-week test is "making a lot fewer much fewer errors," and on refactors it "cleaned up a lot of logs, files, refactored code, took up bits" with "much sketch gaslighting" (his phrasing for hallucinated output). Crucially, M3 "doesn't tend to overthink things anymore" — no visible chain-of-throught sprawl compared to the previous build.

Claude Code Integration

M3 runs inside Claude Code by swapping the base URL to the Anthropic-compatible MiniMax endpoint, so all MCPs, agent teams, and fan-out review features still work. The creator says this Claude Code setup is "the biggest token use" he has — he calls M3 his "daily driver" specifically because of this workflow.

Comparison to Chinese Alternatives

While testing Chinese models after "Fable 5 was banned" (Anthropic's Claude export-controlled release), he ran Kimi 2.7 and GLM. GLM's quality was "pretty good," but the cost "went up quite aggressively" and he "burnt through my 5-hour allowance rather quickly" — in one hour he used ~20% of his MiniMax weekly quota on GLM, which is the wrong way around.

Where M3 Does Not Win

For planning, Opus 4.8, GPT 5.5, and Z are "definitely smarter." The creator uses those for "large functions and planning out," then hands the plan to M3 to "execute, and I can just burn away those tokens, no problem." A 3-video monthly bonus is included but "expensive" per generation, so he plans to batch them in a "column job."

Catch

The hourly token counter is opaque — "they don't tell you exactly how many millions of tokens you used."

Watch on YouTube

Prefer the native player? Open it on YouTube: https://www.youtube.com/watch?v=hTkxebQdtH8