# openai-responses (OpenAI Responses API Examples)

These examples use OpenAI's Responses API. GPT-6 provider IDs such as `openai:gpt-6-sol` select Responses by default.

You can run this example with:

```bash
npx promptfoo@latest init --example openai-responses
cd openai-responses
```

## Examples

### Basic Responses API (`promptfooconfig.yaml`)

Basic example showing how to use the Responses API with GPT-6 Sol, Luna, and Astra, plus a GPT-4.1 comparison model.

### External Response Format (`promptfooconfig.external-format.yaml`)

Compares inline JSON schemas with schemas loaded from `response_format.json` using `file://`. External files let you reuse a schema across configurations.

### Function Calling (`promptfooconfig.function-call.yaml`)

Checks that the response contains a `get_current_weather` function call. This example uses the Responses API function-tool format; it does not execute the weather function.

### Function Callbacks (`promptfooconfig.function-callback.yaml`)

Uses `functionToolCallbacks` to run an `addNumbers` function locally when the model calls it. Assertions check the callback result.

### Reasoning Models (`promptfooconfig.reasoning.yaml`)

Compare GPT-6 Sol, Luna, and Astra using explicit reasoning effort settings.

### GPT-5.1 (`promptfooconfig.gpt-5.1.yaml`)

Example demonstrating GPT-5.1's key features including:

- **`none` reasoning mode**: No reasoning tokens for fastest responses
- **Verbosity control**: Adjustable output length (`low`, `medium`, `high`)
- **Reasoning effort levels**: Compare `none`, `medium`, and `high` reasoning modes
- **Coding tasks**: Optimized for coding and problem-solving workflows

### GPT-5.2 (`promptfooconfig.gpt-5.2.yaml`)

Example comparing GPT-5.2 with different reasoning effort levels:

- **none**: No reasoning tokens for fastest responses
- **medium**: Balanced reasoning for most tasks
- **high**: More reasoning for complex problem-solving

### GPT-5.5 (`promptfooconfig.gpt-5.5.yaml`)

Example comparing GPT-5.5 standard and pro models with different Responses API reasoning settings.

### GPT-5.6 (`promptfooconfig.gpt-5.6.yaml`)

Example comparing the Sol, Terra, and Luna tiers. The `gpt-5.6` alias routes to Sol. All tiers support `max` reasoning; Codex `ultra` is available for Sol and Terra rather than as a Responses API reasoning value.

### GPT-6 Astra (`promptfooconfig.gpt-6-astra.yaml`)

Example using Astra with Responses, `high` reasoning, and structured output. Requires an OpenAI account with Astra access. Astra supports `low`, `medium`, `high`, `xhigh`, and `max` reasoning; tool calling requires Responses. See the [provider documentation](https://www.promptfoo.dev/docs/providers/openai/#gpt-6-astra) for pricing and hosting availability.

### Image Processing (`promptfooconfig.image.yaml`)

Example demonstrating image input capabilities with vision models.

### Web Search (`promptfooconfig.web-search.yaml`)

Example showing web search capabilities.

### Prompt Caching (`promptfooconfig.prompt-cache.yaml`)

Example combining `prompt_cache_key`, `prompt_cache_options`, and included
`web_search_call.results` payloads in a Responses request.

### Codex Models (`promptfooconfig.codex.yaml`)

Example using Codex models for code generation tasks.

### MCP (Model Context Protocol) (`promptfooconfig.mcp.yaml`)

Example demonstrating OpenAI's MCP integration with remote MCP servers. It requires a call to DeepWiki's `ask_wiki_question` tool to query public GitHub repositories. Assertions check that the MCP call succeeds and that the answer includes the expected topic.

#### MCP Features Demonstrated:

- Remote MCP server integration
- Tool filtering with `allowed_tools`
- Approval settings configuration

## Running the Examples

To run any of these examples:

```bash
# Basic Responses API example
npx promptfoo@latest eval -c promptfooconfig.yaml --no-cache

# External response format example
npx promptfoo@latest eval -c promptfooconfig.external-format.yaml --no-cache

# MCP example
npx promptfoo@latest eval -c promptfooconfig.mcp.yaml --no-cache

# Function calling example
npx promptfoo@latest eval -c promptfooconfig.function-call.yaml --no-cache

# Function callbacks example
npx promptfoo@latest eval -c promptfooconfig.function-callback.yaml --no-cache

# Reasoning models example
npx promptfoo@latest eval -c promptfooconfig.reasoning.yaml --no-cache

# GPT-5.1 example
npx promptfoo@latest eval -c promptfooconfig.gpt-5.1.yaml --no-cache

# GPT-5.2 example
npx promptfoo@latest eval -c promptfooconfig.gpt-5.2.yaml --no-cache

# GPT-5.5 example
npx promptfoo@latest eval -c promptfooconfig.gpt-5.5.yaml --no-cache

# GPT-5.6 example
npx promptfoo@latest eval -c promptfooconfig.gpt-5.6.yaml --no-cache

# GPT-6 Astra example
npx promptfoo@latest eval -c promptfooconfig.gpt-6-astra.yaml --no-cache

# Image input example
npx promptfoo@latest eval -c promptfooconfig.image.yaml --no-cache

# Web search example
npx promptfoo@latest eval -c promptfooconfig.web-search.yaml --no-cache

# Prompt caching example
npx promptfoo@latest eval -c promptfooconfig.prompt-cache.yaml --no-cache

```

## Prerequisites

- OpenAI API key set in the `OPENAI_API_KEY` environment variable
- Model access for every configured provider; the basic, function-calling, and reasoning examples include GPT-6 Astra
- For MCP examples: Access to remote MCP servers (some may require authentication)

## Notes

- The MCP example uses the public DeepWiki MCP server which doesn't require authentication
- For production use with MCP, carefully review the data being shared with third-party servers
- Some MCP servers may require API keys or authentication tokens in the `headers` configuration
- External file references support both JSON and YAML formats
- External files are resolved relative to the config file location