MiniMax M3 (free, via OpenRouter)

Zero-cost frontier-ish model on OpenRouter (minimax/minimax-m3:free); ran a real multi-file ADR implementation end-to-end with no visible degradation.

Goals

genops's agent:genops:main session picked up on this model for a stretch of work closing out ADR 0001 — a multi-file operational change on the exhaustive-frame branch plus a research skeleton on exhaustive-frame-orchestrator — with no special babying, just the normal agentic loop: read files, edit config/pipeline manifests, git mv, track a goal via update_goal, ship.

Effectiveness

Good, for this shape of task. The session closed 4 real operational changes and a second branch's research skeleton, called update_goal cleanly to mark the work complete, and drew zero corrections in review. Multi-step tool use (goal tracking, file moves, pipeline.yml edits) held together across the whole session — nothing here required stepping in to unstick a stalled or looping agent.

What made it effective

• The task was well-scoped (an already-written ADR to implement, not open-ended design) — exactly the shape of work where a free model doing mechanical execution is a good trade against paying for a frontier model's judgment.
• Tool discipline looked normal: no sign of hallucinated file paths or skipped verification steps in the reviewed transcript.

Friction / pain points / surprises

This was a single session's worth of evidence, not a stress test — no long-horizon ambiguity, no adversarial edge cases, no large refactor. It's a data point that "free tier, real agentic coding, no degradation" is possible for well-scoped operational work, not a claim that it holds for harder tasks. Worth another look next time a bigger, messier task lands on this model by default (deliberately or via fallback) to see where it actually falls over.