Model Routing

Model Routing

BizarHarness routes every task to the cheapest model that can do it. The dispatch is determined by the agent definition (which model the agent uses) and the routing table in config/AGENTS.md (which agent handles which task type). This page describes the model tiers, the cost-aware dispatch rules, and how to override the model per task.

The 5-tier model architecture

BizarHarness uses a five-tier model hierarchy. Every agent maps to exactly one tier, and the system always prefers the lowest tier that can do the job.

Tier	Cost label	Model	Used by	When to use
0 (free)	Free	DeepSeek V4 Flash (OpenCode Zen)	Frigg, Vör, Quick, Mimir, Heimdall	Read-only Q&A, mechanical edits, codebase search
1 (low)	$	MiniMax M2.7	Hermod, Thor, Baldr	Moderate implementation, git ops, design plans
2 (mid)	$$	MiniMax M3	Odin, Tyr	Complex implementation, architecture, deep debugging
3 (auditor)	$$	MiniMax M3 (edit-deny)	Forseti	Adversarial plan review
4 (last)	$$$$	GPT-5.5 (OpenAI subscription)	Vidarr	Last-resort fallback when all else fails

The escalation path is Free → $ → $$ → $$$$. The default for any non-trivial request is Tier 1 (Thor on M2.7) because it's the cheapest model with the implementation skill to handle moderate work. Tier 2 (Tyr on M3) is reserved for genuinely complex work. Tier 4 (Vidarr on GPT-5.5) is never the default; it's the escape hatch.

Cost-aware dispatch

Odin follows three rules when routing:

1. Default to the cheapest capable agent. A bug fix goes to Thor, not Tyr. A rename goes to Heimdall, not Thor. A clarification goes to Vör, not Odin.
2. Escalate only when justified. A "plan the architecture" request goes to Tyr because the work is large. A "trace the source of the memory leak" request goes to Tyr because the debugging requires M3's reasoning depth. A "postmortem of failed attempts" goes to Vidarr because M2.7 and M3 have both failed.
3. Always parallel for multi-stream work. A request that splits into research + implementation runs Mimir (free) and Thor ($) in parallel, not sequentially. The wall-clock cost is the slowest stream, not the sum.

The cost escalation for a typical request:

"add a /healthz endpoint"
   → Heimdall (free) if it's truly trivial
   → Thor ($)  if it spans a few files
   → Tyr ($$)  if it requires design decisions
   → Vidarr ($$$$)  only if Thor and Tyr both failed

Per-agent model configuration

The model for each agent is set in its Markdown frontmatter. The full list:

Agent	Model string in frontmatter
Odin	minimax/MiniMax-M3
Frigg	opencode/deepseek-v4-flash-free
Vör	opencode/deepseek-v4-flash-free
Quick	opencode/deepseek-v4-flash-free
Mimir	opencode/deepseek-v4-flash-free
Heimdall	opencode/deepseek-v4-flash-free
Hermod	minimax/MiniMax-M2.7
Thor	minimax/MiniMax-M2.7
Baldr	minimax/MiniMax-M2.7
Tyr	minimax/MiniMax-M3
Vidarr	openai/gpt-5.5
Forseti	minimax/MiniMax-M3

The format is <provider>/<model> (e.g., minimax/MiniMax-M3). This is the opencode model identifier. To change an agent's model, edit its .md file in config/agents/ and re-run the installer.

Override model per task

To run a single task with a different model, pass --model to the subagent dispatch (the task tool) when invoking it from Odin. The syntax is opencode-native:

@odin:task --model minimax/MiniMax-M3 research the auth module

In practice, most users don't override — Odin's routing is already cost-aware. The override is for the case where you know the work needs a specific model.

For direct agent invocations, the model is fixed by the agent's frontmatter. If you want a different model for a one-off task, route through Odin with the override.

Adding custom models

To add a new model (e.g., for a different provider), three steps:

1. Connect the provider in opencode. Run /connect in the TUI and add the provider. Verify with /models.

2. Add a model entry in the agent's frontmatter. Edit the target agent's .md file:

   ---
   name: my-agent
   model: my-provider/my-model
   ---

3. Update config/AGENTS.md if the model is in a new tier. The routing table documents which agents use which models. Keep it in sync with the agent files.

Models are matched by the <provider>/<model> string. If the string doesn't match a connected provider, opencode returns an error at dispatch time. Use /models in the TUI to confirm the model is available before assigning it.

Disable a model tier

To disable a tier entirely, remove the agent definitions that use it. For example, to disable the GPT-5.5 tier:

1. Delete or move config/agents/vidarr.md out of the agents directory.
2. Re-run ./install.sh or bizarharness.
3. Restart opencode.

Odin will skip Vidarr in its routing table. The escape hatch is gone — Tier 4 work will simply fail rather than escalate to GPT-5.5.

To temporarily disable Vidarr without removing the file, set its model: to a less expensive model in its frontmatter (e.g., minimax/MiniMax-M3).

Next steps

Next: Bizar Plugin — the bundled opencode plugin that detects subagent loops, reports per-session activity, and injects handoff messages.