Vercel AI Gateway 上 Baseten 决策模型限免到 10.27，价表为 0

Vercel AI Gateway 把 Baseten 的两个决策模型 Mammoth-1、Elephant-1 放到免费档，changelog 写明免费到 2026 年 10 月 27 日。模型名是 baseten/mammoth-1 和 baseten/elephant-1。2026 年 10 月 23 日早上拉了一遍公开模型列表，两条的 pricing.input、pricing.output 都是 "0"。

依据是 Vercel 10 月 21 日的 changelog Baseten's Mammoth-1 and Elephant-1 now available on AI Gateway for free，调用方式写在 Decision 文档里。条款和价表是 10 月 23 日对着 changelog、文档和 GET https://ai-gateway.vercel.sh/v1/models 核的。

限免窗口和模型规格

changelog 原文是 “Free access is available through October 27”，并写了 “for free on AI Gateway for one week”。Mammoth-1 是 27B，Elephant-1 是 9B，上下文都是 16k。它们回答的是 Choice、Boolean、Score 这类带概率的结构化问题，不是自由文本生成。文档里写，决策请求不走 Chat Completions、Responses，也不走 Anthropic、Cohere 兼容接口，要走 POST /v1/evaluate，或者 OpenAI 兼容的 POST /v1/decisions。

当天公开模型列表里，两条相关记录是这样的：

[
  {
    "id": "baseten/mammoth-1",
    "type": "evaluation",
    "pricing": {
      "input": "0",
      "output": "0"
    },
    "context_window": 16384
  },
  {
    "id": "baseten/elephant-1",
    "type": "evaluation",
    "pricing": {
      "input": "0",
      "output": "0"
    },
    "context_window": 16384
  }
]

changelog 举例里用的是 AI SDK 的 decide，HTTP 侧则是 /v1/evaluate，请求体带 model、state、questions。鉴权文档写的是 AI Gateway API Key，或者 Vercel OIDC token，放在 Authorization: Bearer ... 里。

用 Spring Boot 调 /v1/evaluate

Spring Boot 4.1.1、JDK 21，依赖只用 web 和 restclient。Key 从环境变量读：

ai.gateway.api-key=${AI_GATEWAY_API_KEY:}
ai.gateway.model=baseten/mammoth-1
server.port=8080

package com.example.decide;

import java.util.Map;

import org.springframework.beans.factory.annotation.Value;
import org.springframework.http.MediaType;
import org.springframework.stereotype.Component;
import org.springframework.web.client.RestClient;

@Component
public class DecideClient {

    public record EvaluateRequest(String model, String state, Map<String, Question> questions) {}

    public record Question(String type, String instructions) {}

    private final RestClient restClient;
    private final String model;

    public DecideClient(RestClient.Builder builder,
                        @Value("${ai.gateway.api-key}") String apiKey,
                        @Value("${ai.gateway.model}") String model) {
        RestClient.Builder b = builder.baseUrl("https://ai-gateway.vercel.sh");
        if (!apiKey.isBlank()) {
            b = b.defaultHeader("Authorization", "Bearer " + apiKey);
        }
        this.restClient = b.build();
        this.model = model;
    }

    public String evaluate(String state, String instructions) {
        EvaluateRequest body = new EvaluateRequest(
                model,
                state,
                Map.of("refund", new Question("boolean", instructions)));
        return restClient.post()
                .uri("/v1/evaluate")
                .contentType(MediaType.APPLICATION_JSON)
                .body(body)
                .retrieve()
                .body(String.class);
    }
}

控制器把网关返回的状态码和响应体原样透出来：

@RestController
public class DecideController {

    private final DecideClient client;

    public DecideController(DecideClient client) {
        this.client = client;
    }

    @GetMapping("/decide")
    public ResponseEntity<?> decide(
            @RequestParam(defaultValue = "I was charged twice. Please refund the duplicate.") String state) {
        try {
            return ResponseEntity.ok(client.evaluate(state, "Is the customer asking for a refund?"));
        } catch (HttpStatusCodeException e) {
            return ResponseEntity.status(e.getStatusCode()).body(e.getResponseBodyAsString());
        }
    }
}

Maven 打包日志里有 BUILD SUCCESS，结束时间是 2026-10-23T08:05:42+08:00。随后启动到 8080：

2026-10-23T08:05:51.416+08:00  INFO ... o.s.boot.tomcat.TomcatWebServer : Tomcat started on port 8080 (http) with context path '/'
2026-10-23T08:05:51.424+08:00  INFO ... com.example.decide.DecideApplication : Started DecideApplication in 1.351 seconds (process running for 1.793)

这次没有可用的 AI Gateway Key，请求是真发到网关的，但没有拿到决策结果。不设 AI_GATEWAY_API_KEY，以及填一个假 Key，访问本地 /decide 时网关都回：

{"error":{"message":"Authentication failed","param":null,"type":"authentication_error"}}
HTTP 401

文档示例里成功响应大致是这样（取自 Decision 文档，模型名按示例保留）：

{
  "model": "typesafe-ai/jev",
  "answers": { "refund": { "type": "boolean", "probability": 0.98 } },
  "usage": { "inputTokens": 275, "outputTokens": 20 }
}

前几天写过 AI Gateway 上 MiniMax M3.1 Flash 限免，那条走 Chat Completions，而且限免日已经到了。这次是决策模型，接口换成 /v1/evaluate。更早还有 LangChain4j 接 OpenAI Decisions，形状接近，只是模型换成 Baseten 这两条免费档。