Chat Completions
An OpenAI-style single-turn chat completion API format, widely used for integration with existing LLM application code and SDKs.
Core Facts
Timeline (last 90 days)
The article's view: if the use case is concentrated on standard Chat Completions and does not involve realtime connections, the benefits of upgrading to v3.36.0 are relatively limited.
M3.1 Flash can be called via the AI SDK, Chat Completions, Responses, and the Anthropic Messages API, with the model name minimax/minimax-m3.1-flash.
Developers can call Glimpse through the AI SDK or the Chat Completions interface using the model name elorian/glimpse.
The default communication interface of the OpenAI Agents SDK is the Responses API, not Chat Completions.
OpenAI Chat Completions streaming responses do not include usage information by default; only when stream_options: { include_usage: true } is explicitly set will an additional chunk carrying the usage field be sent before the stream ends, and the choices array of that chunk is empty.
In the chunk carrying usage returned by OpenAI Chat Completions streaming, choices is an empty array; if code directly accesses chunk.choices[0].delta.content without optional chaining, it will throw an exception on that chunk.
In the usage field of OpenAI Chat Completions, prompt_tokens includes cached tokens (prompt_tokens_details.cached_tokens is a subset of it); when calculating cost, the cached portion must first be subtracted and full price applied to the remainder.
Switching the same OpenAI Agents SDK Agent from the default Responses API to Chat Completions reduced format-error answers from 41 to 9, and switching to a public grounding system prompt reduced them to 20, indicating that API defaults are the main cause of format drift.
Among the three endpoints of the Muse Wonder API, the reasoning effort parameter is named differently: reasoning.effort in Responses, and top-level reasoning_effort in Chat Completions.
On the Chat Completions endpoint, reasoning content is not persisted across turns, causing inconsistency before and after in multi-turn agent conversations; the fix is to switch to the Responses API.
8 more timeline events
All Facts (19)
OpenAI API格式已成为大模型领域事实上的行业标准接口规范，其核心是/v1/chat/completions端点
80%VerifiedM3.1 Flash can be called via the AI SDK, Chat Completions, Responses, and the Anthropic Messages API, with the model name minimax/minimax-m3.1-flash.
75%UnverifiedThe article's view: if the use case is concentrated on standard Chat Completions and does not involve realtime connections, the benefits of upgrading to v3.36.0 are relatively limited.
50%UnverifiedDevelopers can call Glimpse through the AI SDK or the Chat Completions interface using the model name elorian/glimpse.
50%UnverifiedThe default communication interface of the OpenAI Agents SDK is the Responses API, not Chat Completions.
50%UnverifiedIn the chunk carrying usage returned by OpenAI Chat Completions streaming, choices is an empty array; if code directly accesses chunk.choices[0].delta.content without optional chaining, it will throw an exception on that chunk.
50%UnverifiedOpenAI Chat Completions streaming responses do not include usage information by default; only when stream_options: { include_usage: true } is explicitly set will an additional chunk carrying the usage field be sent before the stream ends, and the choices array of that chunk is empty.
50%UnverifiedIn the usage field of OpenAI Chat Completions, prompt_tokens includes cached tokens (prompt_tokens_details.cached_tokens is a subset of it); when calculating cost, the cached portion must first be subtracted and full price applied to the remainder.
50%UnverifiedSwitching the same OpenAI Agents SDK Agent from the default Responses API to Chat Completions reduced format-error answers from 41 to 9, and switching to a public grounding system prompt reduced them to 20, indicating that API defaults are the main cause of format drift.
50%UnverifiedAmong the three endpoints of the Muse Wonder API, the reasoning effort parameter is named differently: reasoning.effort in Responses, and top-level reasoning_effort in Chat Completions.
50%UnverifiedOn the Chat Completions endpoint, reasoning content is not persisted across turns, causing inconsistency before and after in multi-turn agent conversations; the fix is to switch to the Responses API.
50%UnverifiedChat Completions, Responses and Anthropic Messages are different interfaces, and when a gateway claims OpenAI compatibility, it usually only guarantees part of the behavior of one of these interfaces, which does not mean all three are available.
50%UnverifiedOn Chat Completions, gpt-6-luna's function calling is available only when reasoning effort is none; using built-in tools requires switching to the Responses API.
50%UnverifiedDevelopers can control Gemini 4 Argon's thinking level via the reasoning parameter in the AI SDK or the reasoning_effort parameter in the Chat Completions interface.
50%UnverifiedIn the traditional Chat Completions paradigm, system messages are usually only allowed to appear at the beginning of the message list
50%UnverifiedAny service provider that supports the Chat Completions format is applicable to the fallback chain scheme described in this article.
50%UnverifiedThe none effort parameter has been removed from the schema, and legacy system tool-calling capabilities that relied on this parameter to execute function calls at the chat completions endpoint will stop working and must be migrated to the Responses API.
50%UnverifiedThe Responses protocol is a communication protocol supported by the Microsoft Agent Framework that, compared to the Chat Completions protocol, is better suited for stateful multi-turn conversations and streaming output scenarios.
50%UnverifiedChat mode corresponds to OpenAI's /v1/chat/completions endpoint, taking a list of multi-turn conversation messages as input
50%