OpenAI rate-limit metadata is always EmptyRateLimit in Spring AI 2.0.0

[wonderful-coding-life]

Bug description

When using the OpenAI chat model with Spring AI 2.0.0, the rate-limit metadata returned by ChatResponse is always an instance of EmptyRateLimit.

The completion, model metadata, and token usage are populated correctly, but the OpenAI rate-limit response headers do not appear to be mapped to ChatResponseMetadata.

I believe rate-limit metadata was available in earlier Spring AI 2.0.0 milestone versions, possibly up to 2.0.0-M1, but it is no longer available in the final 2.0.0 release.

Environment

Spring AI: 2.0.0
OpenAI chat model
Model: gpt-5-mini-2025-08-07
Java: 21

Steps to reproduce

ChatResponse chatResponse = chatClient.prompt()
        .user("Which country has the largest population in Europe?")
        .call()
        .chatResponse();

log.info("completion = {}", chatResponse.getResult().getOutput().getText());

log.info("model = {}", chatResponse.getMetadata().getModel());

var usage = chatResponse.getMetadata().getUsage();
log.info("usage = {}", usage);

var rateLimit = chatResponse.getMetadata().getRateLimit();
log.info("rateLimit = {}", rateLimit);

The completion, model, and usage information are returned correctly.

model = gpt-5-mini-2025-08-07
usage = DefaultUsage{
    promptTokens=20,
    completionTokens=655,
    totalTokens=675,
    cacheReadInputTokens=0
}
rateLimit = org.springframework.ai.chat.metadata.EmptyRateLimit@24eecabf

chatResponse.getMetadata().getRateLimit() always returns EmptyRateLimit.

Expected behavior

The OpenAI rate-limit response headers should be mapped to ChatResponseMetadata, allowing request and token limits, remaining values, and reset times to be accessed through getRateLimit().

[cfcromn]

Hi, I’d be interested in working on this once the issue has been triaged.

I can first verify where the OpenAI rate-limit headers are lost—whether they are available from the underlying client response but not mapped into ChatResponseMetadata, or whether the current client abstraction does not expose them.

If the headers are available, I could add the missing mapping and a regression test using a mocked response containing the relevant rate-limit headers.