Which Model Should You Route To?
Map task type to model tier using verified per-million-token input prices. Pick the cheapest tier whose quality is acceptable for the failure cost of the task.
Routing Decision Matrix
Match your workload to the cheapest tier that meets your quality bar.
| Task type | Recommended tier | Rationale |
|---|---|---|
| High-volume classification, routing, extraction | Gemini 2.0 Flash | Cheapest input rate; throughput-friendly for background jobs |
| Chat assistants, JSON-tool agents, general Q&A | GPT-4o Mini | Low-cost tier with mature OpenAI ergonomics and tool-use |
| Long-context reasoning, code review, dense-doc summarization | Claude Haiku 4.5 | Better quality-per-token bracket for tasks where errors are expensive |
Rule of thumb: pick the cheapest tier whose quality is acceptable for the failure cost of the task. If a wrong answer costs you a refund, you don't route to Gemini Flash. If a wrong answer just costs you a retry token, you don't route to Haiku.
Verified input pricing — May 2026
USD per 1M input tokens. These are the figures the calculator uses.
| Model | Provider | Input $/M | Notes |
|---|---|---|---|
| Gemini 2.0 Flash★ Cheapest Google | $0.10 | Lowest-cost tier; verify output rate before committing spend forecast | |
| GPT-4o Mini OpenAI | OpenAI | $0.15 | Lowest-cost tier; verify output rate before committing spend forecast |
| Claude Haiku 4.5 Anthropic | Anthropic | $0.80 | Lowest-cost tier from Anthropic; verify output rate before committing spend forecast |
Cost per Workflow
Pick a workflow to see the input-token cost of a single run on each tier. Output rates are typical (3–5× input); figures below are input-only unless noted, so they're a lower bound on real spend — verify output rates before forecasting.
Cost per task — common workflows
| Workflow | Gemini 2.0 Flash | GPT-4o Mini | Claude Haiku 4.5 | Cheapest |
|---|
Token profiles above are shipped assumptions for a single representative task — batch summarization assumes 100 pages fed in (50k input tokens), real-time chat assumes 8k context plus user message, code generation is a spec + repo-context call, and classification handles one short record per task. Your real numbers will vary — swap to Custom to model your own.
Blended-Cost Calculator (generic)
Estimate monthly input-token cost for your workload across all three tiers. Output cost and tier mixing are out of scope for this stub.
Now track what each route actually costs
Pick a routing strategy, then watch real spend in production. SpendPilot tracks per-model cost, surfaces budget overruns, and catches agent loops before they hit your invoice.
Start tracking free →