Liquid AI's d1 is available on AI Gateway

Liquid AI has added its d1 decision model to Vercel's AI Gateway, giving developers a unified way to run structured classification, routing, and scoring workloads alongside every other model they already access through the gateway. The addition means decision calls made against d1 appear in the same logs and count toward the same cost budgets as standard text generation requests.

This article covers what d1 actually does, how it differs from conventional language models, what vision support unlocks, and how developers can start calling it today through the AI SDK, the OpenAI-compatible Decisions API, or the TypeSafe-compatible API. It is aimed at engineers building classification pipelines, content moderation layers, or any system that needs structured, probabilistic answers rather than free-form generated text.

What shipped in Liquid AI's d1 is available on AI Gateway

Liquid AI's d1 is a decision model, which is a distinct category from a generative language model. Where a standard LLM produces text tokens, d1 evaluates a shared state — a string or JSON payload describing a situation — against a set of typed questions and returns structured answers with probabilities attached. It supports three question types: Boolean, Choice, and Score. A Boolean question asks something like "Was a refund issued?" and returns a true/false answer; a Choice question asks the model to select from a named set of criteria; a Score question returns a numeric rating. None of these paths produce running prose, which keeps latency and cost tied to the decision task rather than to token generation length.

The model is available on AI Gateway under the model identifier liquid/d1. Vercel's AI Gateway is a unified proxy that lets applications call many different model providers through a single endpoint, with shared logging, cost tracking, and budget enforcement. Adding liquid/d1 to that roster means developers can route decision calls through the same infrastructure they use for GPT-4o or Claude calls, without standing up a separate integration or managing a second billing relationship.

What improved in Liquid AI's d1 is available on AI Gateway

The most significant capability extension accompanying this launch is vision support. d1 can now accept image data as part of its state payload, allowing typed questions to be answered about the contents of an image rather than only about text descriptions. Use cases Liquid AI identifies include visual classification, visual inspection, and scoring — meaning the model can, for example, look at a product photograph and return a Choice answer selecting a color, or assign a Score to an image against a rubric, all without generating a caption or any natural-language output. The image is passed as a base64-encoded data URI embedded directly in the state array.

Because the source material does not include numeric benchmark comparisons, latency figures, or pricing tables, a before/after table is not applicable here — no such figures appeared in the changelog. What is concrete is the architectural difference the decision API enforces: rather than a prompt-and-completion round trip, the call specifies state and questions as typed fields, and the SDK returns an answers object whose keys map directly to the question names defined in the call. That schema-first contract eliminates post-processing steps like regex extraction or output parsing that generative calls typically require.

What you gain from Liquid AI's d1 is available on AI Gateway

For engineers running classification or routing pipelines, the practical benefit is that d1 removes the need to prompt-engineer a generative model into returning a consistent schema and then validate that schema on every response. The answers object returned by a d1 call is typed at the question level: a Boolean question yields a boolean, a Choice question yields one of the named criteria, a Score question yields a number. That predictability matters in production systems where a malformed output from a generative model can break a downstream routing decision or trigger a retry loop.

Vision support extends this reliability to image pipelines. A team building a content moderation system, for example, can pass flagged images directly into a d1 state payload and ask typed questions about what the image contains, receiving a structured answer with probabilities rather than a written description that must then be parsed. Similarly, an e-commerce team can ask classification questions about product images — color, category, defect presence — without maintaining separate fine-tuned classifiers. Because these calls flow through AI Gateway, usage appears in the same dashboard as all other model calls, making cost attribution straightforward.

How to get Liquid AI's d1 is available on AI Gateway

Access to liquid/d1 runs through Vercel's AI Gateway using one of three integration paths. The first is the AI SDK's experimental_decide function. The second is the OpenAI-compatible Decisions API. The third is the TypeSafe-compatible API. All three paths use liquid/d1 as the model string. The AI SDK path looks like this:

```typescript
import { experimental_decide as decide } from 'ai';

For image input, the state field accepts a JSON array containing a question string and an image object with a base64 data URI. The Choice question type requires a criteria object whose keys name the options — { red: null, blue: null, green: null } for a color classification, for instance. There is no separate installation step beyond having the AI SDK and an AI Gateway endpoint configured; the model identifier liquid/d1 routes the call automatically. Vercel's decision documentation covers Choice, Score, and Boolean question specs in full, and a decision model guide is available alongside a browsable model listing in the gateway.

What to watch after Liquid AI's d1 is available on AI Gateway

The decision model category remains early-stage, and d1's placement on AI Gateway is the first integration of this type on the platform. Whether Vercel adds more decision-oriented models from other providers — or whether Liquid AI expands d1's question types beyond Boolean, Choice, and Score — will determine how broadly the pattern gets adopted. The experimental_decide function name in the AI SDK signals that the API contract is not yet stable, which means teams building on it should expect the interface to change before it reaches a non-experimental export.

Longer term, the convergence of decision calls and generative calls under one gateway roof raises the question of compound workflows: a pipeline that uses d1 to classify an input, then routes to a generative model only when the score crosses a threshold, all billed and logged in a single place. That architecture is now technically viable with this launch, though the AI SDK's experimental status suggests the ergonomics are still being worked out. Developers who want to track the roadmap can follow Liquid AI's decision model guides and the AI Gateway changelog for updates on stable API releases and any new question types.