ChatOpenAI (Responses API, streaming): response_metadata.model_name is the requested alias, not the resolved model

[SiKreuz]

Checked other resources

• This is a bug, not a usage question. For questions, please use the LangChain Forum (https://forum.langchain.com/).
• I added a very descriptive title to this issue.
• I searched the LangChain.js documentation with the integrated search.
• I used the GitHub search to find a similar question and didn't find it.
• I am sure that this is a bug in LangChain.js rather than my code.
• The bug is not resolved by updating to the latest stable version of LangChain (or the specific integration package).

Example Code

No API key needed — the converter is exported from the package root:

import { convertResponsesDeltaToChatGenerationChunk } from "@langchain/openai";
import { concat } from "@langchain/core/utils/stream";

// The two events around a streamed Responses call: `response.created` echoes
// the requested model, `response.completed` carries the resolved snapshot.
const events = [
  { type: "response.created", response: { id: "resp_1", model: "gpt-5.4" } },
  {
    type: "response.completed",
    response: {
      id: "resp_1",
      model: "gpt-5.4-2026-03-05",
      status: "completed",
      output: [],
    },
  },
];

const merged = events
  .map((e) => convertResponsesDeltaToChatGenerationChunk(e))
  .filter((c) => c != null)
  .map((c) => c.message)
  .reduce((a, b) => concat(a, b));

console.log("model_name:", merged.response_metadata.model_name);
console.log("model:     ", merged.response_metadata.model);

Actual output on @langchain/openai 1.5.10:

model_name: gpt-5.4
model:      gpt-5.4gpt-5.4-2026-03-05

Expected: both gpt-5.4-2026-03-05. The non-streaming path already does this —
convertResponsesMessageToAIMessage allowlists model: response.model and
model_name: response.model from the completed response.

### Error Message and Stack Trace (if applicable)

_No response_

### Description

When the Responses API is used with streaming, `response_metadata.model_name` is the **requested** model string rather than the resolved snapshot the API actually served, and `response_metadata.model` is a concatenation of both.

In `convertResponsesDeltaToChatGenerationChunk` (`src/converters/responses.ts`), both keys are written from the `response.created` event, which echoes back the requested model:

```typescript
} else if (event.type === "response.created") {
  id = event.response.id;
  response_metadata.id = event.response.id;
  response_metadata.model_name = event.response.model; // requested alias
  response_metadata.model = event.response.model;
}

The response.completed branch carries the resolved snapshot, but only copies the keys present on its response object:

for (const [key, value] of Object.entries(event.response)) { ... }

event.response has no model_name key, so model_name is never corrected. model is written a second time — and because neither model nor model_name appears in the last-write-wins list _mergeDicts consults (["id", "name", "output_version", "model_provider"]) nor in DEFAULT_MERGE_IGNORE_KEYS, the second write is string-appended instead of replacing:

model_name: "gpt-5.4"
model:      "gpt-5.4gpt-5.4-2026-03-05"

Expected: both "gpt-5.4-2026-03-05", matching the non-streaming Responses path (convertResponsesMessageToAIMessage, which sets model_name: response.model) and the Completions path.

Impact: anything attributing cost or usage by response_metadata.model_name records the alias, so per-snapshot pricing and model-version comparisons are wrong. Langfuse's LangChain callback handler ends every generation with this field, so all traced reasoning-enabled calls are recorded under the alias.

Note model_provider already sits in the merge's last-write-wins list, so this class of key is handled — model and model_name look simply missed. #11054 is the same merge behaviour surfacing on the Completions path.

Proposed fix

Write both keys from response.completed / response.incomplete instead of response.created:

  } else if (event.type === "response.created") {
    id = event.response.id;
    response_metadata.id = event.response.id;
-   response_metadata.model_name = event.response.model;
-   response_metadata.model = event.response.model;
  } else if (event.type === "response.completed" || event.type === "response.incomplete") {
    id = event.response.id;
+   response_metadata.model_name = event.response.model;

(model is then set once by the existing Object.entries copy, so nothing concatenates.) Adding model and model_name next to model_provider in _mergeDicts' last-write-wins list would additionally fix #11054 and make this robust for every provider.

Related

• ChatOpenAI streaming duplicates response_metadata.model_name / finish_reason (and doubles response_metadata.usage) when the provider emits multiple finish_reason chunks (e.g. OpenRouter) #11054 / PR fix(openai): keep scalar response_metadata last-wins across finish_reason chunks #11084 — the same merge-concat class on the Completions path. fix(openai): keep scalar response_metadata last-wins across finish_reason chunks #11084 fixes it with a local mergeScalarResponseMetadata in ChatOpenAICompletions._streamResponseChunks, so it does not reach the Responses converter.

• ChatOpenRouter streaming responses do not preserve routed model metadata #11409 — same theme (streaming loses the real model) in @langchain/openrouter.

• ChatOpenAI (Responses API, streaming): response_metadata embeds full request tools schemas on every AIMessage #11162 / PR fix(openai): stop persisting echoed request fields in streamed response_metadata #11163 — please read together with this issue. fix(openai): stop persisting echoed request fields in streamed response_metadata #11163 replaces the Object.entries(event.response) copy with Object.entries(msg.response_metadata). That allowlist contains both model: response.model and model_name: response.model, while response.created still writes both keys with the requested alias — so after fix(openai): stop persisting echoed request fields in streamed response_metadata #11163 the merge concatenates model_name too:

  today:        model_name: "gpt-5.4"                     model: "gpt-5.4gpt-5.4-2026-03-05"
  after #11163: model_name: "gpt-5.4gpt-5.4-2026-03-05"   model: "gpt-5.4gpt-5.4-2026-03-05"
  

  Removing the two response.created writes (below) is the prerequisite that makes fix(openai): stop persisting echoed request fields in streamed response_metadata #11163 safe.

System Info

• @langchain/openai 1.5.10, @langchain/core 1.2.9, langchain 1.5.10 (all latest at time of filing)
• macOS (Darwin arm64)
• Node.js 26.7.0 / Bun 1.4.0