Tiers

# Tiers

Low → high. `frontier` is the most capable/expensive rung; `balanced` is the fast-and-capable middle
(formerly the "quick" tier).

| Tier | OpenRouter model | Context | $ in / out per 1M | Role |
|---|---|---:|---|---|
| **Budget** | `deepseek/deepseek-v4-flash-0731` | 1,048,576 | $0.09 / $0.18 | Cheapest. High-volume, cost-dominated work. |
| **Balanced** | `google/gemini-2.5-flash` | 1,048,576 | $0.30 / $2.50 | Fast + capable everyday middle; interactive/structured work. |
| **Performance** | `google/gemini-3.7-flash` | 1,048,576 | $0.375 / $1.875 | Deepest non-frontier reasoning; builds the durable substrate. |
| **Frontier** | `moonshotai/kimi-k3` | 1,048,576 | $3 / $15 | Kimi K3 — 2.8T-param reasoning model. Where quality compounds. |

> Retuning a whole rung (swap the model) is a one-line change in `TIER_MODEL`; every category on that
> tier follows. `google/gemini-2.5-flash-lite` is the cross-model `FALLBACK_MODEL` (last-resort
> safety net after provider failover), not a tier; its context window is also 1,048,576 tokens.
> The executable limits live in `MODEL_CONTEXT_WINDOWS`, sourced from OpenRouter's routed-provider
> catalog values. Generation preflight reserves output and reasoning room before admitting a call.

Raw Markdown source