Public accessAI SDK 7.0.59 completed five bounded live Responses requests on August 11, 2026.Read current status
by Flow7
Sign inCreate account

Infer × Vercel AI SDK · live check @ai-sdk/openai 4.0.37

Point the OpenAI Responses provider at Infer.

Create a direct provider with Infer’s /v1 base URL, then select .responses(...). Every captured SDK call stayed on /v1/responses.

Use the checked setupRead the receiptsCheck the selector
clientai 7.0.59
provider@ai-sdk/openai 4.0.37
schemazod 4.4.3
runtimeNode 23.10.0
request path/v1/responses
selectorinfer/gpt-5.6-terra:balanced

Pinned setup

Install the versions that were checked.

The base URL ends at /v1; the explicit Responses model factory appends /responses. Keep the Infer key in the environment.

npm install --save-exact ai@7.0.59 @ai-sdk/openai@4.0.37 zod@4.4.3
Before the first paid call

Read current status for infer/gpt-5.6-terra:balanced, verify the account email, fund the wallet, and create a capped API key.

maxOutputTokens: 64 was the proof ceiling, not a general output-policy recommendation. The validation-only fetch observer recorded paths and assigned a fresh idempotency key to each outbound call; it is not needed to select the Responses endpoint.

Checked configuration@ai-sdk/openai · Responses
import { createOpenAI } from "@ai-sdk/openai";
import { generateText } from "ai";

const infer = createOpenAI({
  name: "infer",
  apiKey: process.env.INFER_API_KEY,
  baseURL: "https://infer.flow7.org/v1",
});

const model = infer.responses(
  "infer/gpt-5.6-terra:balanced",
);

const { text } = await generateText({
  model,
  prompt: "Your bounded task",
  maxOutputTokens: 64,
  maxRetries: 0,
  telemetry: { isEnabled: false },
  providerOptions: {
    openai: {
      store: false,
      strictJsonSchema: true,
    },
  },
});

console.log(text);

This code makes a direct SDK-to-Infer request. It does not use Vercel AI Gateway. Check status and test the workload before raising output or spend limits.

Dated production record

Five completions. Five receipts. $0.004222 total.

Checked August 11, 2026 with retries disabled and two isolated API keys, each capped at exactly $0.05 daily and monthly. Both keys were revoked after request, receipt, wallet, and reservation reconciliation. This is a compatibility record, not workload or uptime certification.

TextgenerateText returned the exact bounded text for $0.000464: rcpt_5f70276e8753797b9317e3a8.
SDK SSEstreamText consumed the Responses event sequence and exact text for $0.000462: rcpt_56b2916c812b4ff1e1c42ae7.
Tool callA forced call to one local read-only function completed for $0.001496: rcpt_46aee7172ffe29f9f7495ccc.
Tool resultThe next SDK step sent function_call and function_call_output, then returned the exact nonce for $0.000270: rcpt_21764e11574fb67f0d9d7b48.
Output.objectThe request sent text.format.type = "json_schema" and returned the validated object for $0.001530: rcpt_1710163bf79046d7edee6d1f.
Serializer diagnosticsBefore the final matrix, two completed responses were billed but rejected by the SDK: a message item needed an id and array-valued annotations (rcpt_93b75a017a658445de0d5633, $0.000710), then a function_call item needed an id and status (rcpt_153f8acb1a084f762da16b7c, $0.001402). Infer normalized both item types before the five-request matrix passed.
Restart interruptionOne separate request was interrupted by a host restart. It has no receipt, a $0.000000 customer charge, and a fully released reservation. No billable supplier evidence was recorded; this does not prove an external call never began.
Full validation ledgerThe successful matrix cost $0.004222; the two paid diagnostics cost $0.002112; the interrupted request cost $0.000000. Cumulative customer charge: $0.006334.
Wire pathEvery captured Vercel AI SDK request posted to /v1/responses. No /v1/chat/completions request was observed.

Claim boundary

What this proves—and what it does not.

  1. 01

    AI SDK, not Vercel AI Gateway.

    The checked setup sends requests directly to Infer through @ai-sdk/openai. It does not use or validate Vercel AI Gateway, and it is not evidence of a Vercel partnership, endorsement, or certification.

  2. 02

    A custom provider, not a native Infer package.

    Infer is configured with createOpenAI and a custom base URL; it is not a named provider built into AI SDK. @ai-sdk/openai-compatible was not used because the current package targets Chat Completions rather than this Responses proof.

  3. 03

    SSE-compatible, not real-time upstream streaming.

    streamText consumed Infer’s Responses SSE sequence. Infer currently buffers the upstream response before emitting that sequence, so this record does not establish first-token latency or long-stream behavior.

  4. 04

    One narrow local tool loop.

    The proof forced one read-only local function, executed it once, and completed one tool-result turn. Parallel calls, state-changing tools, approvals, repair, MCP, agent classes, and built-in or hosted tools remain untested.

  5. 05

    One native schema path.

    The checked Output.object request used one strict Zod schema mapped to Responses-native json_schema. It does not certify arbitrary schemas, multimodal input, files, audio, or images.

  6. 06

    No state, UI, or recovery claim.

    previousResponseId, Conversations, background mode, cancellation, retries, failover, AI SDK UI, framework hooks, and long-running workloads remain untested. Retries were set to zero.

  7. 07

    Versions and selector are bounded.

    The record covers the pinned Node and package versions with infer/gpt-5.6-terra:balanced. Other selectors, tiers, releases, and broader Responses features require their own checks.

  8. 08

    No supplier attestation.

    The receipts prove Infer-routed requests completed and record customer charges. They do not disclose or attest the upstream supplier identity.

Primary referencesOfficial source and API documentation checked August 11, 2026.
Responses POST source ↗Schema mapping source ↗Responses options source ↗Responses API ↗Streaming events ↗

Start with one capped request.

Create an account, verify the email, add service credit, and check the selected route before running the pinned setup.

Create accountRead the Responses contract