Assign different models to different tasks. Haiku searches your codebase, Sonnet writes the code, Flash compacts context.

Open the router

/router

Or edit ~/.empryo/config.json:

{
  "taskRouter": {
    "spark":      "anthropic/claude-haiku-4-5",
    "ember":      "anthropic/claude-sonnet-4-5",
    "explore":    "anthropic/claude-haiku-4-5",
    "webSearch":  "anthropic/claude-haiku-4-5",
    "desloppify": "anthropic/claude-haiku-4-5",
    "verify":     "anthropic/claude-haiku-4-5",
    "goalReview": "anthropic/claude-sonnet-4-5",
    "review":     "anthropic/claude-sonnet-4-5",
    "compact":    "google/gemini-2.5-flash",
    "semantic":   "anthropic/claude-haiku-4-5",
    "default":    null
  }
}

What each slot does

Slot	Runs when	Good choice
`spark`	Dispatched read-only agents explore your code	Fast/cheap (Haiku, Flash)
`ember`	Dispatched agents that edit files	Strong coding model (Sonnet, Opus)
`explore`	The solo `explore` agent. Unset falls back to `spark`	Fast/cheap
`webSearch`	Web search agents. Unset disables the agentic path	Fast/cheap
`desloppify`	Cleanup pass after code edits	Fast/cheap
`verify`	Adversarial review after code edits	Medium strength
`goalReview`	The goal-loop reviewer, with its own clean context	Medium/strong
`review`	On-demand `/review`. Unset falls back to `goalReview`	Medium/strong
`compact`	Context compaction	Fast/cheap (Flash is ideal)
`semantic`	Genome one-line symbol summaries	Fast/cheap
`default`	Fallback when no slot matches	-

Leaving a slot unset is a real choice, not just a blank. spark inherits the model your turn is already using, which means it shares that model's prompt-cache prefix — routing it to something cheaper can cost you more in cache misses than it saves per token. Measure before you split it. A slot pointing at a model that fails to resolve degrades to inherit rather than killing the turn.

Recommended setups

Strong code, cheap everything else.

"spark": "anthropic/claude-haiku-4-5",
"ember": "anthropic/claude-sonnet-4-5",
"compact": "google/gemini-2.5-flash"

Why this saves money

Agents spend ~70% of their tokens on exploration (reading files, running greps, navigating the Genome). That work doesn't need an expensive model. Reserve your strong model for the ~30% that's actually writing code.