Parallel Decisions Lands on AI Gateway: Free Access to Real-Time Web Decision Models

Parallel Decisions Fast lands on AI Gateway, offering free structured real-time judgments with probabilities and evidence.
Parallel's decision model Parallel Decisions Fast is now callable through AI Gateway, free until Nov 23, 2026, with no separate Parallel account required — just use parallel/parallel-decisions-fast. Instead of generating conversation, it searches the live web and returns probabilities and cited evidence for Boolean, Choice, and Score questions, with related questions in one request sharing context and evidence. Citations apply to the whole request rather than per answer. Threshold-based automation with human escalation suits CRM enrichment, eligibility checks, lead routing, and claim verification, and calls are logged and budgeted through AI Gateway.
Developers can now call Parallel's decision model, Parallel Decisions Fast, for free through AI Gateway — no separate Parallel account needed. Just use the model name parallel/parallel-decisions-fast. The model targets dynamic information about companies, products, policies, and more, searching the live web in real time and returning structured answers with probabilities and cited evidence that applications can act on directly.

What Is Parallel Decisions Fast
Parallel Decisions Fast is a decision model from Parallel. It isn't designed to generate conversational text — its goal is to answer factual questions whose answers change over time. The model searches the live web and returns probability values and cited evidence for the questions you ask.
This solves a common pain point: traditional LLMs have knowledge cut off at their training data, so for questions like "does this company offer an official CLI" or "is this policy currently in effect," they tend to give outdated or speculative answers. Parallel Decisions anchors answers to currently verifiable sources through real-time retrieval and provides a confidence probability.
On AI Gateway, the model is free to use until November 23, 2026, and does not require a separate Parallel account.
Training knowledge cutoffs vs. retrieval-augmented generation (RAG): General-purpose LLMs have knowledge frozen at their training data cutoff and cannot perceive changes after that point, so for time-sensitive questions they can only "guess" from parametric memory. Retrieval-augmented generation (RAG), by contrast, retrieves the latest information from external sources at inference time, then has the model answer based on the retrieved content. Parallel Decisions follows the retrieval-augmented approach, but its output isn't natural-language paragraphs — it's structured fields with probabilities. The retrieval pipeline typically includes query rewriting, web page fetching, snippet extraction, and relevance ranking, with the final evidence returned to the caller as citations so answers can be traced and verified.
Three Question Types and Shared Context
Parallel Decisions Fast supports three structured question formats covering different decision scenarios:
- Boolean questions: Return the probability of "yes." Suitable for eligibility checks, feature verification, and other binary scenarios.
- Choice questions: Pick one answer from options you define. Suitable for classification routing and lead assignment.
- Score questions: Evaluate and score against an ordered rubric you provide. Suitable for risk grading and priority ranking.
A single request can include multiple related questions, which share context and retrieved evidence. This means retrieval costs are amortized and consistency across related judgments is better. Official typical uses include enriching CRM records, checking eligibility criteria, routing sales leads, and verifying claims against current sources.
Probabilistic binary classification and threshold decisions: Boolean questions return the model's confidence probability of "yes" (between 0 and 1) rather than a definite true/false. This lets downstream systems use thresholds to convert continuous probabilities into discrete actions — for example, auto-writing to the CRM above 0.9, flagging for review between 0.6 and 0.9, and routing straight to a human below 0.6. Choice and Score questions work the same way: you can set acceptance conditions on the option distribution or scoring results. Probability calibration quality determines how well threshold strategies work. In practice, evaluate the model's accuracy and confidence distribution on a batch of samples with known answers across question types before setting thresholds.
Calling the Model: Three Interface Forms
The official examples show the same web-retrieval question asked through the AI SDK, the TypeSafe-compatible API, and the OpenAI-compatible Decisions API. Here's the AI SDK example using the experimental_decide function:
import { experimental_decide as decide } from 'ai';
const result = await decide({
model: 'parallel/parallel-decisions-fast',
state: {
company: 'Clay',
website: 'https://www.clay.com'
},
questions: {
has_cli: {
type: 'boolean',
instructions: 'Does this company offer an official command-line interface?'
}
},
providerOptions: {
parallel: { includeExcerpts: true }
}
});
console.log(result.answers.has_cli);
console.log(result.providerMetadata?.parallel?.citations);
The state field passes in known context (such as company name and website), and the model retrieves based on it. Enabling includeExcerpts attaches evidence excerpts to the results.
AI SDK and OpenAI-compatible interfaces: AI SDK is Vercel's unified model-calling library that uses provider abstraction to let the same code switch between model providers. The experimental_ prefix indicates the API is still experimental and its signature may change. OpenAI-compatible interfaces follow the OpenAI Chat Completions request format, so existing code built on the OpenAI SDK only needs its base URL and model name swapped to migrate, reducing rework. The TypeSafe-compatible API emphasizes typed request and response definitions, enabling compile-time checks in TypeScript projects. All three call the same backend model; the differences are mainly in request wrapping and response structure.
Cited Evidence and Probability-Driven Automation
Citation information includes source URLs, optional titles, and evidence excerpts. Note that citations apply to the entire request rather than mapping each source to an individual answer. In other words, the model returns a set of evidence supporting the request, and the application must determine which evidence is relevant to each specific question.
Probability values are designed for automation workflows: applications can set thresholds to automatically act when answer probability is above the threshold, while lower-confidence results go to human review. This "auto-execute plus escalate uncertainty to humans" pattern is better suited to business process integration than having an LLM output free-form text.
Integration Value with AI Gateway
When you call the model through AI Gateway, requests are logged and counted against spending budgets alongside other model requests, making observability and cost management consistent. For teams already using AI Gateway, adopting decision models requires no new account system or billing channel — integration cost is very low.
In terms of product positioning, Parallel Decisions fills the capability layer of "structured judgments that require factual grounding": retrieval fetches current evidence, the model outputs probabilities and options consumable by programs, and business logic makes the decision. For high-frequency judgment tasks with volatile answers — lead routing, eligibility verification, claim checking — this is a better engineering fit than general-purpose chat models. Developers can view details in the model catalog and refer to the decision model documentation for specifics on each question type.
The free period is a good window to evaluate whether this type of model fits your business — especially for judgment steps that have long relied on manual verification and are hard to cover with rules.
AI Gateway's role: AI Gateway is a unified proxy layer between applications and multiple model providers, handling authentication, routing, retries, rate limiting, logging, and usage billing. Developers only need Gateway credentials to call models from different vendors using a unified model naming scheme (e.g., parallel/parallel-decisions-fast) without opening separate accounts for each. Request logs and spending budgets are presented centrally, making it easy to attribute costs by model or team and perform observability analysis. The value for enterprises is reduced operational burden from multi-vendor integration while retaining flexibility for future model swaps or A/B comparisons.
One-line Summary
Parallel Decisions Fast is now on AI Gateway, offering free structured real-time judgments with probabilities and evidence.
Paragraph Summary
Parallel's decision model Parallel Decisions Fast is now available through AI Gateway, free until November 23, 2026, with no separate Parallel account required — just use the model name parallel/parallel-decisions-fast. Rather than generating conversation, the model retrieves the live web in real time and returns probability values and cited evidence for Boolean, Choice, and Score questions, with multiple related questions in a single request sharing context and retrieval results. Citations include source URLs, titles, and excerpts, but apply to the entire request rather than mapping to individual answers. Applications can set probability thresholds to take automated actions, with low-confidence results escalated to human review — ideal for CRM enrichment, eligibility verification, lead routing, and claim checking. Calls through AI Gateway are also integrated into unified logging and spending budget management.
Related articles

Gemini 4 Neon Released: A Balance Between Frontier Intelligence and Efficient Task Execution
Google DeepMind releases Gemini 4 Neon, focused on frontier intelligence and fast end-to-end task execution. This article breaks down its benchmark performance in coding, enterprise analysis, scientific reasoning, and cybersecurity, its four thinking levels, pricing, and safety alignment data.

Is AI the Culprit Behind Global Outages? The Trust Crisis Behind the NSA and FAA Failures
Fourth major failure in five days hits NSA classified network. Microsoft blames a software bug, FAA denies its new AI system, NSA says routine maintenance. Senator Warner suggests AI accelerates vulnerability discovery, leading to rushed fixes — while 39% of Americans already view AI negatively, exposing a deepening trust crisis.

Robot Developer Builds Custom B-rep Kernel with 3,487 Tests
A robot developer builds a custom B-rep geometry kernel with 3,487 tests, exploring why robotics needs dedicated CAD tools and the barriers to building one.