Teapot Router / Documentation
Choosing a model: the Claude and GPT families
Read the model names and tiers on both lanes, pick the right model per task, and stop overpaying.
Open workspaceWhat the two lanes are
- Teapot Router keeps exactly two model lanes: the Claude lane (Anthropic protocol) and the GPT lane (OpenAI protocol).
- A lane is a protocol family, not a capability tier — each lane spans flagship to lightweight models.
- One account and one key work on both lanes; each request is billed by the model it actually used.
For the full live model list and prices, open the pricing ledger on the site.
Reading the Claude family
- Common tiers: Opus is the flagship — deepest reasoning for architecture-scale refactors and hard problems; Sonnet is the balanced daily driver; Haiku is the fast, cheap tier for bulk and small tasks.
- Tiers get version bumps over time; newer versions are usually more capable and may be re-priced.
- Higher tier, higher price: heavy work on Opus, daily work on Sonnet, chores on Haiku is the standard cost-control combo.
| Tier | Name means | Reasoning profile | Best for | Price |
|---|---|---|---|---|
| Opus | A grand musical work | Deepest: multi-step reasoning, long-horizon tasks, cross-file understanding | Architecture refactors, stubborn bugs, deep analysis | Highest |
| Sonnet | A short, crafted poem | Balanced: enough for almost all engineering work | Daily coding, code Q&A, docs | Mid |
| Haiku | A minimal three-line poem | Light and fast: simple calls at low latency | Completions, classification, bulk edits | Lowest |
Exact model IDs and current prices live in the pricing ledger.
Reading the GPT family
- The GPT lane also splits by generation and suffix: flagship models for complex work, lighter models tuned for speed and cost.
- The number marks the generation; suffixes usually flag a specialty or speed tier — OpenAI's notes are the source of truth.
- The selection logic matches Claude: flagship for deep reasoning, lightweight for high-volume calls.
| Tier | Naming pattern | Reasoning profile | Best for | Price |
|---|---|---|---|---|
| Flagship | The mainline model with the highest generation number | Deepest: complex multi-step reasoning, long-horizon tasks | Architecture, stubborn bugs, deep analysis | Highest |
| Balanced | The standard model of a mainline generation | Balanced: enough for daily engineering work | Daily dev, code Q&A, docs | Mid |
| Lightweight / special | Suffixed variants (mini, fast style) | Light or specialized: low latency, tuned for a scenario | Completions, bulk processing, special tasks | Lower |
Don't guess capability from a name; weigh the ledger price against your task.
Picking by task
- Architecture, cross-file refactors, stubborn bugs: flagship tier (Claude Opus or a GPT flagship).
- Everyday features, code Q&A, docs: the balanced tier (e.g. Claude Sonnet) is the value pick.
- Completions, bulk edits, formatting scripts: a lightweight tier (e.g. Haiku) is plenty and cheap.
- Unsure? Run the balanced tier once, then adjust based on duration and cost in Usage.
| Task type | Recommended tier | Claude side | GPT side |
|---|---|---|---|
| Architecture / cross-file refactor | Flagship | Opus | Flagship model |
| Daily dev / code Q&A | Balanced | Sonnet | Balanced model |
| Completion / bulk processing | Lightweight | Haiku | Lightweight model |
| Not sure | Start balanced, adjust | Sonnet | Balanced model |
Flagship is not a universal upgrade: on simple tasks it just costs more and runs slower.
Where to check prices and cost
- The site's pricing ledger lists input and output rates for every model on both lanes.
- The workspace Models catalog shows the IDs your key can actually use — copy from there when configuring a client.
- Each request's model, tokens, cost, and duration is in Usage; use it to verify a model choice pays off.
The price table sets expectations; Usage is the final bill.