Baseten Mammoth-1 and Elephant-1 Decision Models Launch Free on AI Gateway for a Limited Time

AI Gateway recently announced that two decision-specific models from Baseten, Mammoth-1 and Elephant-1, are now available for limited-time free access on its platform. For developers building automated decision workflows, this offers a low-cost opportunity to experiment.

What Are Decision Models?

Unlike common generative large models, Mammoth-1 and Elephant-1 are Decision Models. They do not generate free-form text; instead, they perform classification, routing, and scoring tasks on input content. You can think of them as programmable judgment engines: given a piece of text and a set of predefined questions, the model returns options, probabilities, and scores.

This design fills the gap between generative models and rule engines. For scenarios such as customer support ticket triage, content moderation, and risk scoring, calling a decision model directly is often more controllable and easier to integrate than letting a general-purpose large model "freestyle."

AI Gateway is a unified model access layer launched by Vercel. Developers can call models from multiple providers through a single API key without separately integrating each provider's SDK and billing system. It handles routing, logging, rate limiting, and usage tracking along the request chain, and is natively compatible with Vercel's AI SDK. For the scenarios in this article, the value of AI Gateway lies in consolidating Baseten's decision models and existing generative model calls into the same interface and billing view, so teams don't need to maintain two sets of integration code when mixing the two types of models.

Jev Decision Index and Artificial Analysis Decision Index are third-party leaderboards designed for decision-type tasks. Their evaluations typically cover accuracy and consistency in intent classification, entity-extraction judgments, multi-label scoring, and similar dimensions—not generation quality or general knowledge Q&A. Rankings on such leaderboards reflect a model's relative performance on structured judgment tasks and are not necessarily positively correlated with parameter count, so it's not contradictory that the 9B Elephant-1 ties with larger models. Note that leaderboard scores are based on public test sets; actual performance on your own business data still needs to be verified.

Specifications and Benchmark Performance of the Two Models

The core difference between the two models is scale:

• Mammoth-1: 27B parameters, the larger of the two. According to official information, it ranks first on the Jev Decision Index.
• Elephant-1: 9B parameters, more lightweight. It ties for first place with other models on the Artificial Analysis Decision Index.

Both support a 16k token context window, and results include selected options, probabilities, and scores. This means developers can ask multiple questions about the same input in a single request—for example, simultaneously determining "is a refund being requested," "which support team should handle this," and "how urgent is it."

One detail: the model itself only provides probabilities; the actual automation threshold is decided at the application layer—you can set "if probability exceeds 0.9, handle automatically; otherwise, route to human review." This design, which separates judgment from execution, makes the decision process more transparent.

How to Call It in Code

AI Gateway provides a unified API entry point, integrated with the AI SDK. The following example demonstrates a refund intent judgment scenario:

import { experimental_decide as decide } from 'ai';

const result = await decide({
  model: 'baseten/mammoth-1',
  state: 'I was charged twice. Please refund the duplicate.',
  questions: {
    refund: {
      type: 'boolean',
      instructions: 'Is the customer asking for a refund?',
    },
  },
});

console.log(result.answers.refund);

Developers simply specify the model name as baseten/mammoth-1 or baseten/elephant-1. The platform also offers integration methods such as the Decisions API and TypeSafe-compat; for specific question types and usage, refer to the decision model documentation.

The experimental_decide prefix indicates that this interface is still experimental. The AI SDK does not guarantee semantic version compatibility for such APIs, and signatures may change when upgrading minor versions. Before integrating into production, lock dependency versions and prepare a fallback plan. In the example, questions is an object keyed by question name; each question uses type to declare the expected output form (such as boolean, enum, or numeric), and instructions to provide a natural-language description. The returned result.answers returns structured results under the same keys, including the selected option and its corresponding probability. This convention of "declarative questions + structured responses" allows model outputs to be consumed directly by the type system and downstream business logic, eliminating the need to write regex or JSON repair code to parse free-form text.

Free Window and Usage Management

This free access has a clear time window, ending on October 27. After the free period ends, decision calls will appear in AI Gateway logs like other model requests and will be counted toward budget statistics.

For teams evaluating whether decision models suit their business, this one-week free period is enough to run prototype validation. It's recommended to prioritize testing two types of scenarios: first, classification tasks that require high accuracy but have fixed output structures; second, complex routing logic that requires joint judgment across multiple questions.

A Brief Observation

The decision model category is attracting more attention. Although general-purpose large models are capable, they often suffer from unstable output and high cost when it comes to "making judgments." Baseten's choice to use specialized small models for such tasks is similar to using dedicated chips for specific computations—trading generality for determinism and efficiency. Mammoth-1 and Elephant-1 cover effectiveness-first and cost-first needs respectively, and this product stratification is a fairly pragmatic choice for developers.

The free period is an observation window; the real test is whether, after it ends, these models' performance under real business traffic lives up to their leaderboard rankings.