# Zabio

> Zabio is an OpenAI compatible LLM router for models from Anthropic, OpenAI, Google, DeepSeek, xAI, Mistral AI, Together AI and TypeSafe AI, under one key and one prepaid balance. Send the model `auto` and Zabio names the job, picks the model and reasoning setting that fit it on the line you choose, can split a long-document request across models when that is cheaper, and reports what every call cost and what it saved. Every call pays the provider list price plus Zabio's service fee, a percentage of the provider price, and every `auto` request also pays a small fixed fee for automatic model choice, once per request. Zabio takes no cut of what routing saves, nor of a cache or batch discount: every cent of the saving stays with the customer. Calls on the customer's own vendor keys pay no service fee; their `auto` requests pay the fixed fee. The fees in force are on https://zabio.ai/pricing/ and, as JSON, at https://api.zabio.ai/v1/public/plan.

Key facts:

- Base URL: https://api.zabio.ai/v1. OpenAI compatible: the official OpenAI SDKs for Python and Node, and tools built on them such as LangChain and the Vercel AI SDK, work unchanged; change the base URL and the API key.
- Catalog: eight vendors, the chat models of Anthropic, OpenAI, Google, DeepSeek, xAI, Mistral AI and Together AI, and TypeSafe AI's Jev decision model. Every model with its price: https://zabio.ai/models/
- Routing: with plain `auto` (or no model), Zabio picks for you: the cheapest model able to do the job, moving up a line when the job needs it. `auto/best`, `auto/balanced` and `auto/cheap` set the trade-off; any model can be named directly. A classifier (TypeSafe AI's Jev, about 0.15 seconds a call, with Anthropic's Haiku as fallback) names the task, a difficulty from 1 to 5, the stakes and the language. Picks come from public benchmark results by task and difficulty. High-stakes jobs move up a line. On easy and medium jobs the cheap and balanced lines use the model's lowest reasoning setting. Failover moves to the next model by the same rule, up to three attempts.
- Smart Splitting: under `auto`, a request that carries a long document and asks about part of it can be split. A cheap model copies out the lines the question needs, word for word, Jev checks the extract holds what the question needs, and the model auto picked answers from the excerpt. It happens only when it is cheaper, never on a named model, every step shows in `zabio.split`, and `zabio.split: "off"` turns it off. A failed attempt costs the customer nothing. In Zabio's tests it cut the cost of long-document questions by about a third, with every fact kept.
- Jev for classification: name `jev-1.13.0` with a JSON schema made of fixed choices, yes or no answers and bounded numbers, and get the JSON with a calibrated probability per field in `zabio.decision`. Under `auto`, strict-schema classification jobs go to Jev whenever it is the cheaper answer.
- Caching: for Anthropic, Zabio marks the system prompt, tools and conversation history for the vendor's cache; OpenAI, DeepSeek, xAI and Mistral cache on their own. When a session's turns are more than five minutes apart, Zabio uses the 1 hour cache. Cached tokens are billed at the vendor's cached rate.
- Flex: `zabio.deadline` of 15 minutes or more lets Zabio use the flex tier of OpenAI or Google at about half the price; a capacity refusal is retried at the normal tier and billed at the price used.
- Batching (since 1 October 2026): `/v1/files` and `/v1/batches` in the format of OpenAI's Batch API; every line routed like a normal call at the vendors' batch prices (about half the list price at OpenAI, Anthropic, Google, Mistral and Together AI; Jev and Smart Splitting work inside batches, and a batch can opt in to run lines without a batch tier at the normal price), answers within 24 hours; inside a batch `auto` only picks models with a batch tier, and a line naming a model without one stops the batch before anything is billed. Up to 50,000 lines and 100 MiB per batch. No file needed for single requests (since v1.2.1-rc28): a Responses API request with background: true and a zabio.deadline of 24 hours or more joins a batch pool that goes out every 15 minutes. https://zabio.ai/features/batching/
- Pricing: prepaid, in US dollars. Each call costs the provider list price plus the service fee (a percentage of the provider price), and each `auto` request also pays a fee for automatic model choice, once per request whatever happens inside it: the model's own cost for the request, at least the plan's `auto_call_fee_floor_usd` and at most its `auto_call_fee_usd` (a flat `auto_call_fee_usd` on a plan without a floor), so a small request pays less. The classifier is never billed on its own; it is part of what that fee pays for. Zabio takes no share of the saving. Every fee and the minimum top-up come from a versioned pricing plan: the plan in force, and any scheduled change, is at https://api.zabio.ai/v1/public/plan (JSON, no key) and on https://zabio.ai/pricing/. Prices can change, and the pricing page always shows the current ones. The saving each response reports is measured against the key's baseline: the 95th percentile of the prices of its calls over the last 30 days, priced with the same cache hits the call had. A new key needs 20 calls before its baseline is set. A named model is its own baseline, so it shows no saving. Every account starts on a Free plan: 100 requests a day on six models from four vendors (OpenAI, Google, Mistral, Together AI), up to 16,000 tokens in and 2,000 out per request, no card, no expiry; the first top-up removes every limit.
- Operator: Zabio is operated by THINGS & STUFF, LLC, a Delaware limited liability company.
- Feature support by vendor and model (tools, images, structured output, reasoning, prompt caching, flex, batch), built from the live catalog: https://zabio.ai/models/feature-support/
- Official channels: YouTube https://www.youtube.com/@zabio-ai and X https://x.com/zabioai
- Articles: https://zabio.ai/articles/ (essays and guides by the people who build Zabio; the first is "Why I built Zabio", the founder's story: juggling models, keys and a growing AI bill, then building a router tuned to save you money)
- Launch offers (since 29 September 2026, open to every organization, testers included): the match, where an organization billed at least 250 USD in the 30 days after its first top-up gets every later top-up matched dollar for dollar up to 250 USD; and referrals, 25 USD of credit for each new organization a user brings that adds 100 USD in top-ups, paid once those top-ups are 30 days old. Offer credit joins the balance, works on any call, never expires and has no cash value. Details: https://zabio.ai/referrals/ and https://zabio.ai/pricing/#earn-credit
- Teams: unlimited organizations and API keys, owners, admins and members, a monthly quota per member, per-key spend caps, rate limits and allowed models, an audit log, and your own vendor keys with no service fee.
- Privacy: prompts are not stored unless body logging is on or a stored conversation is used; stored content is encrypted per organization.
- Status: beta. Models, prices and features can change.
- Languages: English (https://zabio.ai/) and Spanish (https://zabio.ai/es/).

## Pages

- [Pricing](https://zabio.ai/pricing/): the fees in force, a worked example, how the saving is measured, credits and the fine print
- [Features](https://zabio.ai/features/): [Smart Splitting](https://zabio.ai/features/smart-splitting/), [smart routing](https://zabio.ai/features/smart-routing/), [Jev for classification](https://zabio.ai/features/jev/), [caching](https://zabio.ai/features/caching/), [flex and batching](https://zabio.ai/features/flex/), [teams and keys](https://zabio.ai/features/teams-and-keys/)
- [How Zabio compares](https://zabio.ai/vs/): a dated, sourced comparison with [OpenRouter](https://zabio.ai/vs/openrouter/), [LiteLLM](https://zabio.ai/vs/litellm/) and [Portkey](https://zabio.ai/vs/portkey/)
- [Questions and answers](https://zabio.ai/faq/)
- [Calculator](https://zabio.ai/calculator/): estimate a month's bill through Zabio from your own volume, token sizes and kind of work, with Zabio's fees included

## Docs

- [Quickstart](https://zabio.ai/docs/core/quickstart/): base URL, API key, first call with curl, Python and Node
- [Models and routing](https://zabio.ai/docs/core/models/): the auto lines, how routing picks a model, reasoning settings, prices
- [Chat completions](https://zabio.ai/docs/core/chat-completions/): the OpenAI compatible chat endpoint, Smart Splitting, deadlines and Zabio's extra fields
- [Responses API and conversations](https://zabio.ai/docs/core/responses/): stateful conversations across models and vendors
- [Embeddings](https://zabio.ai/docs/core/embeddings/)
- [Errors](https://zabio.ai/docs/core/errors/): error codes and what to do about them
- [API reference](https://zabio.ai/docs/core/api-reference/)
- [Command line](https://zabio.ai/docs/core/cli/)

## Account and billing

- [Billing](https://zabio.ai/docs/cloud/billing/): credits, the service fee, the fee for automatic model choice and worked examples
- [Your own vendor keys](https://zabio.ai/docs/cloud/own-keys/): bring your own key for any vendor Zabio serves, no service fee
- [Organizations, members and keys](https://zabio.ai/docs/cloud/teams/)
- [Privacy and data](https://zabio.ai/docs/cloud/privacy/)
- [Roadmap](https://zabio.ai/docs/cloud/roadmap/)

## Optional

- [Models](https://zabio.ai/models/): every model with its price, context window and capabilities
- [News](https://zabio.ai/news/): model releases, price changes and vendor announcements, with what each one means for calls through Zabio. RSS: https://zabio.ai/news/feed.xml
- [Status](https://zabio.ai/status/): live health of the API, the vendor connections and the web app. The same data as JSON: https://api.zabio.ai/v1/status
- [Documentación en español](https://zabio.ai/es/docs/core/quickstart/)
- Contact: hello@zabio.ai