Comparing Minimax 2.5 vs Claude Opus in OpenClaw

Minimax 2.5 felt like 60–70% of Opus, not the 95% the marketing claimed — and the missing 30% is what made the daily briefing land at 8:47 a.m. instead of 7:00.

## The two bots Boxmining ran a split test inside OpenClaw: bot Stark runs Claude Opus on Anthropic's paid plan, bot Jeff runs Miniax 2.5. The boss gave Stark the two agents that actually had to ship; Jeff got the cheap side. The cost gap is the headline number — $30 to $60 per day on Opus for the multi-agent setup, versus $20/month on the Miniax Pro plan. Run that out: 30 × $60 = $1,800/month on Opus.

## What failed on Jeff The 7:00 a.m. cron job for a daily news briefing never fired. After the prompt was rewritten the day before, the new schedule "didn't register it as a cron job." The day Jeff did deliver, the response walked the user to the car wash to wash the car — "decided to send you walking to the car wash without your car." The same car-wash prompt hit Opus correctly on the first try.

## Where Opus actually slipped Stark isn't flawless either. On Feb 26 the daily briefing header was stamped Feb 24; the body content was correct for the 26th. So even at $60/day you get a date-stamp bug. But the report format, the slide spec, and the proactive "in case you missed it" follow-up all hit — the "bonus work" the cheap model never attempted.

## The slot-machine problem The same prompt gives a different answer on different runs. The car-wash question was correct on the first few Miniax attempts, wrong later. The creator's rule: "ask the same question again" — design the website 6 times with the same prompt and pick the best. Cheap inference makes brute-forcing a slot machine the only reliable workflow.

## Context window is the next cliff A commenter named Note reports Miniax works well at low context but "the moment you go above the 120k context window it feels like I'm talking to chat 3.5," with the intelligence drain described as exponential.