The Reflection API’s OpenAI-compatible endpoint, https://api.reflection.ai/openai/v1, implements the OpenAI Chat Completions and Models endpoints, and no others. Code that uses only those endpoints, through the OpenAI SDKs or a framework built on them, works after you change three settings:
Switch from OpenAI
To move code from the OpenAI API, point the client at the Reflection API and change the model:
Many tools also read the OPENAI_BASE_URL and OPENAI_API_KEY environment variables. A tool that uses only Chat Completions and Models can often switch without code changes once you set them and choose a Reflection model:
A tool that calls another endpoint, such as Responses, won’t work.
Supported endpoints
Supported features
- Text messages with
system, developer, user, assistant, and tool roles
- Streaming, including
stream_options.include_usage
- Tool calling, with
tool_choice and parallel_tool_calls
- Structured outputs with
json_object and json_schema
temperature, top_p, frequency_penalty, presence_penalty, max_completion_tokens, and max_tokens (deprecated)
reasoning_effort
logprobs and top_logprobs, on models that support them
safety_identifier, a stable identifier of up to 128 characters for your application’s end user. Requests are attributed to it when detecting and handling misuse, and the model never sees it. Send a hash of a username or email address rather than the value itself.
Differences
Parameters with fixed values
These parameters are accepted only at the value shown. Any other value returns a 400 error with code unsupported_value.
Parameters without effect
These parameters are accepted but have no effect:
Reasoning effort
Each model accepts only the reasoning_effort values listed in its reasoning.supported_efforts, and any other value returns a 400 error with code unsupported_value. Beam-1.0-501B-A23B accepts low, medium, high, xhigh, and max, and when a request omits reasoning_effort or sets it to null, it applies the default that the models endpoints report in reasoning.default_effort. It always reasons, so there is no value such as none that turns reasoning off. See Reasoning.
Log probabilities
Only models that list max_top_logprobs on the models endpoints return log probabilities. Other models accept logprobs only as false or null, and top_logprobs only as null.
On a model that lists it:
- With
logprobs set to true, each choice’s logprobs.content scores the tokens of message.content, and each streamed chunk scores its delta.content in choices[].logprobs. Reasoning tokens are never scored.
top_logprobs can be set only when logprobs is true, and accepts at most the model’s max_top_logprobs.
- With
logprobs set to true, the output budget, reasoning tokens included, is at most 8,192 tokens. A larger max_completion_tokens or max_tokens is rejected, and when the request sets no limit, the default applies capped at 8,192.
Other parameters
Parameters that aren’t listed in the API reference aren’t supported and return a 400 error with code unsupported_parameter. If a framework sends one of these by default, turn it off in the framework’s settings.
Inside a message, a text content part, or a tool_choice object, fields that aren’t listed are accepted and ignored, as in OpenAI’s API, so clients that add fields of their own to earlier messages work unchanged. developer messages are the exception: they refuse unlisted fields. A field whose name differs from a listed one only in case, such as Content, is refused everywhere.
Message content is text only: a string, or an array of text parts. Image, audio, and file inputs are not supported.
Additions
The API returns some fields that OpenAI’s API does not:
Errors
Errors use the same {"error": {"message", "type", "param", "code"}} shape, so SDK exception handling works unchanged. Some code values are specific to Reflection; see Errors. Last modified on November 18, 2026