---
meta:
title: Choosing an API — Meta Model API
description: Compare the Responses, Chat Completions, and Messages formats on Meta Model API — same models, same auth, same cost — and pick the one your code already speaks.
keywords: Meta Model API, Responses API, Chat Completions, Messages API, OpenAI compatible, Anthropic compatible, choosing an API
cms:
layout: large
alias: /model-api/docs/protocols
target: aidmc
---
# Choosing an API
Model API speaks three request formats — the [Responses API](/docs/protocols/responses), [Chat Completions](/docs/protocols/chat-completions), and the [Messages API](/docs/protocols/messages). All three run the same models, use the same bearer-token auth, and cost the same per token. Pick the format your code already speaks; the model and features underneath are identical.
## At a glance {#at-a-glance}
| Format | Endpoint | Best for | Reasoning across turns |
| --- | --- | --- | --- |
| **Responses** | `POST /v1/responses` | Agents and multi-step tool loops; the full feature set | Yes — encrypted replay or `previous_response_id` |
| **Chat Completions** | `POST /v1/chat/completions` | Dropping into existing OpenAI messages-array code | No (reasoning is not carried between turns) |
| **Messages** | `POST /v1/messages` | Anthropic-format tools such as Claude Code | Yes — via Anthropic `thinking` replay |
## Responses API {#responses}
The recommended default for new work. It is OpenAI-compatible and exposes the full feature set: cross-turn reasoning replay (stateless encrypted items or server-managed `previous_response_id`), [search grounding](/docs/search-grounding), [tool search](/docs/tool-search), background execution, and [file inputs](/docs/file-handling). Reach for it for agentic and coding workloads where reasoning continuity across tool turns matters. Use the OpenAI SDK pointed at `https://api.meta.ai/v1`. See the [Responses API guide](/docs/protocols/responses).
## Chat Completions {#chat-completions}
The classic OpenAI messages-array endpoint. It is the simplest drop-in if your code already calls `/v1/chat/completions`: point the client at the base URL, swap the key, and keep the rest. It does not carry reasoning across turns for external keys, and [search grounding](/docs/search-grounding) runs on the Responses API only. [Tool calling](/docs/tool-calling) with `function` tools is fully supported. See the [Chat Completions guide](/docs/protocols/chat-completions).
## Messages API {#messages}
An Anthropic Messages-compatible wire format over the same inference pipeline as Responses. Use it for Claude-oriented tools and the Anthropic SDK — point the SDK at the base host `https://api.meta.ai` (it appends `/v1/messages`) with your `MODEL_API_KEY`. It is stateless, so replay history client-side. See the [Messages API guide](/docs/protocols/messages).
## How to choose {#how-to-choose}
- **Building an agent or coding workflow**: use the **Responses API** for reasoning continuity, search grounding, and background execution.
- **Already have OpenAI `chat.completions` code**: start with **Chat Completions** for the fastest drop-in, and move to Responses when you need reasoning replay or search grounding.
- **Using Claude Code or the Anthropic SDK**: use the **Messages API** and keep your existing Anthropic-format requests.
## Next steps
- Make your first call in the [Quickstart](/docs/quickstart).
- Wire the model into your editor with the [coding agents guide](/docs/coding-agents).
- Verify request and response schemas in the [API reference](/docs/api-reference).
Choosing an API
Model API speaks three request formats — the Responses API, Chat Completions, and the Messages API. All three run the same models, use the same bearer-token auth, and cost the same per token. Pick the format your code already speaks; the model and features underneath are identical.Format | Endpoint | Best for | Reasoning across turns |
|---|
Responses | POST /v1/responses
| Agents and multi-step tool loops; the full feature set | Yes — encrypted replay or previous_response_id |
Chat Completions | POST /v1/chat/completions
| Dropping into existing OpenAI messages-array code | No (reasoning is not carried between turns) |
Messages | POST /v1/messages
| Anthropic-format tools such as Claude Code | Yes — via Anthropic thinking replay |
previous_response_id), search grounding, tool search, background execution, and file inputs. Reach for it for agentic and coding workloads where reasoning continuity across tool turns matters. Use the OpenAI SDK pointed at https://api.meta.ai/v1. See the Responses API guide.The classic OpenAI messages-array endpoint. It is the simplest drop-in if your code already calls /v1/chat/completions: point the client at the base URL, swap the key, and keep the rest. It does not carry reasoning across turns for external keys, and search grounding runs on the Responses API only. Tool calling with function tools is fully supported. See the Chat Completions guide.An Anthropic Messages-compatible wire format over the same inference pipeline as Responses. Use it for Claude-oriented tools and the Anthropic SDK — point the SDK at the base host https://api.meta.ai (it appends /v1/messages) with your MODEL_API_KEY. It is stateless, so replay history client-side. See the Messages API guide.- •Building an agent or coding workflow: use the Responses API for reasoning continuity, search grounding, and background execution.
- •Already have OpenAI
chat.completions code: start with Chat Completions for the fastest drop-in, and move to Responses when you need reasoning replay or search grounding. - •Using Claude Code or the Anthropic SDK: use the Messages API and keep your existing Anthropic-format requests.