Methodology
The classifier matches task descriptions against an open ruleset. It looks for force patterns, signal verbs, signal phrases and exclusion phrases, then chooses between routine, moderate and deep reasoning tiers.
The classifier is not perfect. It handles most clear cases without an AI call. If the confidence score stays below 0.6, the home page offers deep mode for an explicit second pass.
Tier definitions
Deep reasoning
Architecture decisions, security review, complex debugging, large codebase analysis, novel problems.
- Multi-hop reasoning is likely required.
- Output quality is meaningfully higher at larger model tiers.
- Cost difference is justified.
Moderate
Feature work, refactors, debugging known errors, integrations. Multi-step but bounded.
- Some reasoning and synthesis are required.
- Quality improves with stronger coding ability, but frontier depth is not necessary.
Routine
Bounded, well-defined single-step work. Formatting, string and data ops, boilerplate, simple lookups.
- Single-domain knowledge required.
- Output is well-defined and bounded.
- No multi-hop reasoning required.
What the tool does not see
- Large codebases unless that context is included in the task description
- Attached files, full prompt templates, or system instructions
- Provider-specific limits outside the covered model list
Pricing data
Pricing is refreshed nightly from the OpenRouter model catalog and bundled into the static build. The current snapshot was retrieved on April 17, 2026. Costs are shown in USD and estimated from average prompt and completion sizes for each task tier.
The default Any lane uses a curated shortlist from Anthropic, Google and OpenAI. Broader provider coverage still appears in the catalog, but the default recommendation avoids sending most users to a niche provider just because it is the cheapest line item in the cache.
The provider choice is visible before the first recommendation. Many developers already have access to one or two model families, so the first answer should respect that constraint rather than asking them to correct it after the fact.
What does not happen at runtime
- No OpenRouter pricing calls on the heuristic path
- No Gemini call unless you explicitly opt into deep analysis
- No stored history or browser-side session cache
Wrong recommendation?
Open an issue with the task description, the recommendation you saw and the model you think should have been selected. Signal patterns are reviewed manually before merging.