One model id. Each tool call checked.
Send tirtha/auto. With tirtha/auto, Tīrtha picks the model for each task by price, speed and accuracy, checks each tool call against your schema before serving it, and moves up a level when the check fails.
Tīrtha is in public beta. Keys and limits may change. Billing follows the prices listed.
A tool call through the router
Your tools go in the request as usual. The router checks the call the model makes against the schema you sent.
To use one level for every request, pin it: "tirtha": {"mode": "value"} for Base, "quality" for Plus, "max" for Max. Long agent sessions (4+ tool turns or 24K+ input tokens) pinned to Max are served and billed at Plus. Details in the API reference.
The right model answers, then it is checked
- 1
The right model answers
Tīrtha picks the model for each task by price, speed and accuracy.
- 2
Then it is checked
Each tool call is checked against your schema. Code and prose are not tested; the response's label says what was checked.
- 3
Pass: served. Fail: up a level
With tirtha/auto, a request that fails moves up to Plus, then Max, and its tool calls are checked again. On auto, long agent sessions move up to Plus at most.
Keep Claude or Codex for the thinking. Send the building to Tīrtha.
On a Claude or Codex plan: save plan headroom
Send routine tool calls and code generation to Tīrtha. You hit your Claude or Codex limit less often, and the strong model stays available for the hard parts.
Paying per token: save dollars
With tirtha/auto, you pay for Plus or Max only when the check fails.
Levels route to listed models, each under its own licence. Set up a coding agent.
Claude, Claude Code, Codex and other product names are trademarks of their owners; Tīrtha is not affiliated with or endorsed by them.
tirtha/auto, by level.
With tirtha/auto a request starts at Base and moves up only when a tool call fails the check. Prices per 1M tokens, read live from billing. Every model's own price is on Models. About the bill.
What a request costs right now.
Per 1M tokens, read live from the same price table that bills you. Each request pays the price in force when it arrives. Served by vetted US/EU inference hosts.
Base, pinned: $0.14 in / $0.44 out per 1M tokens. Every tool call is checked against your tool schema. If a call fails, Base retries it once with the error. If it still fails, you get the call back marked as failing the schema check, so your code can decide what to do. If a pinned Base tool call fails the schema check, we retry it once on Base, and you pay the pinned Base price for the tokens of both attempts.
If a pinned tool call fails the schema check, Tīrtha removes it from the response. The response still comes back with HTTP 200: you get any text the model wrote, it shows checked: false, and the tokens used, including the failed call, are billed at the pinned price.
If a pinned tool call fails the schema check, Tīrtha removes it from the response. The response still comes back with HTTP 200: you get any text the model wrote, it shows checked: false, and the tokens used, including the failed call, are billed at the pinned price.
Long agent sessions (4+ tool turns or 24K+ input tokens) pinned to Max are served and billed at Plus.
Model names are trademarks of their owners. Tirtha is not affiliated with or endorsed by them.
Prices as of 2026-11-19 22:07 UTC.
Let it choose, or pin a level.
Five modes. auto picks the level for each request. A pin fixes a key or a request to one level. A pinned price is read from the same price table that bills you, shown in the table below.
| Mode | Input | Cached input | Output |
|---|---|---|---|
| auto, served at Base | $0.14 | $0.072 | $0.44 |
| auto, served at Plus | $0.266 | $0.018 | $0.756 |
| auto, served at Max | $2.472 | $0.317 | $7.376 |
| value, pinned to Base | $0.14 | $0.072 | $0.44 |
| quality, pinned to Plus | $0.266 | $0.018 | $0.756 |
| max, pinned to Max | $2.472 | $0.317 | $7.376 |
| frontier | Not available yet | ||
value
Always Base: fast everyday calls.
quality
Always Plus: long agent sessions and multi-step work.
max
Always Max: the most accurate single calls, for when a mistake is expensive.
Long agent sessions (4+ tool turns or 24K+ input tokens) pinned to Max are served and billed at Plus.
All prices are per 1M tokens. Cached input is billed at a lower rate that depends on which host served the request, and it is never above that mode's input price. The exact rate is recorded with each request's bill.
Long agent sessions are never served by Max. If a key pinned to max runs a long session (4 or more tool steps, or 24k or more input tokens), it is served at Plus and billed at Plus.
When auto moves a request up a level, the rejected Base draft is billed too, at Base rates.
When auto moves a request up a level, you pay the level that answered. If a request moves up a level, its first Base attempt is also billed, at Base rates, even if that attempt returned an empty or unusable answer. Every model's own price is on Models.
What it checks, and what it does not
It checks
- Each tool call is checked against your schema: the tool is one you offered and its arguments fit the schema you sent.
- With tirtha/auto, a call that fails moves the request up a level.
- On a pinned level, a failing call is handled as router pricing describes for that level.
It does not
- Code and prose are not tested; the response's label says what was checked.
- A call that fits your schema can still be the wrong call for the task. The check confirms its shape, not its judgement.
- Tīrtha does not run code you send.
Every response says what was checked
A tirtha object sits beside the standard OpenAI fields.
| Field | Meaning |
|---|---|
served_by | Which path answered: the Base model, frontier-value (Plus) or frontier-max (Max), or cache-exact. |
checked, checked_by | checked: true means every tool call in this response passed the schema check; false means a call failed it or no check applied. checked_by names the check that passed, for example schema, and is null when checked is false. |
schema_check | passed or failed on every tool-call answer from tirtha/auto or a pinned level; null when no schema check ran (for example a text or chat answer). |
escalated | true only if the request moved up a level (with tirtha/auto). A pinned level reads false. |
When things fail
In auto mode, a failed check is never marked as passed. If our tests fail or time out, the request moves up to a stronger level and tries again. If no level's answer passes, you still get an answer, marked checked: false. If every level is down, you get a 503, not a guess. Every answer tells you whether it was checked.
Pinned levels stay where you put them. A request pinned to one level is never moved. If its answer fails a check, it comes back marked checked: false. A tool call on Base that fails its format check gets one retry with the error first. If a tool's schema is too unusual for our checker to read, the call is checked for shape only.
We never run your tools. Tool calls come back to you to run. Tīrtha never executes them, so nothing Tīrtha retries can repeat an action in your systems. When Tīrtha tests code a model wrote, it runs it in a locked sandbox with no network access. If a reused connection closes before any reply arrives, we resend the request once on a fresh connection, and you are never charged twice for it.
Your record. Your account shows your balance, your credits and your newest 20 billing entries, plus your recent requests: the level that served each one, whether it moved up, the tokens and the cost. Billing entries are append-only. An entry is never edited or deleted; a correction is a new entry. Our request log keeps counts and labels, never your prompts or answers. How the answer cache works is in our Privacy Policy.
Limits on each key. A key can carry a daily spend cap, a rate limit and a monthly quota. Ask us and we'll set them for you.
Capabilities
For every path through the router: what good means on that path, whether a real check decided it or nothing did, what we measured, and whether that measurement still describes the system serving your request. The page is generated from the configuration we run.