MiniMax · updated 2026-08-29
Run MiniMax M3: providers, pricing and flat-rate access
MiniMax's open-weight M3 — a cheap, capable model served by over a dozen providers, popular as the high-volume workhorse tier in agent stacks.
M3 is one of the cheapest serious models to run: open weights mean 13+ providers compete on price (OpenRouter lists it around $0.23/M input, $0.96/M output; the cheapest hosts blend under $0.25/M). Light users should just pay per token. Standard Compute includes M3 in its flat-rate routed pool for the heavy-usage case.
Your options, honestly
MiniMax APIPay-per-tokenFirst-party access; competitive rates, you manage keys and spend yourself.
OpenRouter / DeepInfra / NovitaPay-per-token13+ hosts on the open weights — real price competition (roughly $0.25–0.75/M input across providers) and fallback options; speed varies a lot by host.
Standard Compute$39+/mo flatM3 in the routed pool as a high-volume workhorse with frontier models behind it — heavy days cost the same as quiet ones. No pinning.
Self-host (vLLM)GPU + powerOpen weights make it real — if you have the hardware and the ops appetite.
the flat-rate trade-off
Honestly: M3's per-token price is low enough that an M3-only workload rarely justifies a subscription. The flat-rate case is mixed workloads — M3-class volume plus frontier-class hard steps in the same agent, where the frontier steps are what blow up a per-token bill.
Running it in your agent
Every OpenAI-compatible agent (OpenCode, Cline, Aider, Continue, Roo) runs M3 with a base-URL swap. On Standard Compute the router sends routine steps to M3-class models automatically.
FAQ
Which MiniMax M3 provider is fastest?
It varies widely — benchmark sites show a ~3x spread in throughput and large latency differences across the 13+ hosts serving the open weights. If you pin one host per-token, check independent benchmarks; routed plans handle provider selection for you.
Is MiniMax M3 good enough for coding agents?
For routine steps — edits, tool use, boilerplate — yes, and at its price that covers a lot of ground. For hard multi-file reasoning, frontier models still lead, which is why M3 works best as the volume tier in a mixed setup.
More: how flat-rate LLM APIs work · per-agent guides for OpenCode, Claude Code and Cline · model comparisons