Published: 2026-04-09
When I set out to configure OpenCode with a multi-model stack, I hit a wall with wallet. The planning agent was Claude Opus 4.6, and that pricing was brutal. We're talking $3 per million input tokens and $25 per million output. For a tool I'd use daily? Not sustainable.
So I went hunting for alternatives.
The mistake most people make when picking models is staring at benchmark leaderboards and calling it a day. But here's the thing: a model that crushes GPQA Diamond isn't necessarily good at orchestrating other agents.
What matters for planning and orchestration is the agentic score — how well the model can reason about tool use, delegation, and coordination. Raw intelligence helps, but it's not the whole story.
I narrowed it down to three models available on OpenRouter that could handle my workload without breaking the bank:
MiMo-V2-Pro (Xiaomi)
MiniMax M2.7
Gemma 4 31B (Google)
I also looked at DeepSeek V3.2 and GLM 5.1, but both had dealbreakers. GLM 5.1 doesn't support tool use at all. So that was a non-starter for agents that need to run commands and delegate work. DeepSeek V3.2 has tool support and is dirt cheap, but the research showed "cracks in search, tools, and complex agent workflows". Which was exactly what I'd be asking it to do.
MiMo-V2-Pro won by being boringly good at the thing I actually needed.
That 82.8% agentic score is the highest of the bunch. The 1.05M context window means it can hold entire project states in memory without breaking a sweat. And it's optimized specifically for agentic scenarios (OpenRouter's words, not mine).
But the real kicker? It's the cheapest of the three viable options at $1/$3 per million tokens. Compare that to Opus at $3/$25 and we're talking massive savings on inference costs.
Here's what I ended up with:
| Agent | Model | Role |
|---|---|---|
| plan | MiMo-V2-Pro | Strategic planning with massive context |
| build | MiniMax M2.7 | Orchestrator — delegates to workers |
| general | Gemma 4 31B | Worker — writes code, edits files, runs commands |
| explore | Gemma 4 31B | Read-only codebase explorer |
| review | MiniMax M2.7 | Post-phase code review |
| qa | MiniMax M2.7 | Pre-commit adversarial testing |
I originally started with Kimi K2.5 for the orchestrator, but MiniMax beat it on every metric that matters: higher intelligence (49.6% vs 46.8%), better coding (41.9% vs 39.5%), higher agentic score (61.5% vs 58.9%), and 3x faster throughput (33 tok/s vs 8 tok/s). The only loss was 57K context window, but 205K is plenty for tracking subagent state.
The real win is the cost structure. Gemma handles the high-volume worker tasks at roughly $0.14/$0.40 per million tokens. Which is effectively free compared to the frontier models. MiniMax covers the quality control where reasoning matters most. And MiMo anchors the planning layer with that unbeatable 1.05M context window.
One thing that surprised me: Gemma's latency is 10.7 seconds with 6.0 tok/s throughput. That's slow. But for worker tasks like editing files and running bash commands, the latency gets hidden behind I/O anyway. You're waiting for the filesystem or network, not the model.
For planning and orchestration where you're waiting on the model to think, MiMo and MiniMax's ~3x speed advantage is noticeable. Nobody wants to stare at a spinner while the orchestrator decides what to do next.
Don't pick models based on hype. Pick them based on the specific capabilities your task requires.
For agent orchestration, agentic score beats raw intelligence. Tool support beats theoretical reasoning. Context window size determines whether your orchestrator can remember what it's doing. And speed determines whether you'll actually enjoy using the tool.
The right stack isn't the most expensive one. It's the one where each model does the specific job you hired it for. While at a price that doesn't make you wince every time you run it.
Compare my final stack to Opus at $3/$25 per million: I'm paying $1/$3 for planning, $0.30/$1.20 for orchestration and QA, and effectively nothing for workers. That's not just cheaper. It's actually sustainable for daily use.
Happy coding,
J.R.