Can't access response HTTP headers

[kevindew]

Hi, one of the features we seem to be missing in this library is the means to access the response headers that were returned with the original HTTP response - in particular to allow us to track rate limit usage for Anthropic First Party and latency information for AWS Bedrock.

For example in a curl request to Anthropic useful information is returned:

anthropic-ratelimit-input-tokens-limit: 10000
anthropic-ratelimit-input-tokens-remaining: 8000
anthropic-ratelimit-input-tokens-reset: 2025-07-23T15:05:10Z
anthropic-ratelimit-output-tokens-limit: 4000
anthropic-ratelimit-output-tokens-remaining: 4000
anthropic-ratelimit-output-tokens-reset: 2025-07-23T15:04:59Z
anthropic-ratelimit-requests-limit: 5
anthropic-ratelimit-requests-remaining: 4
anthropic-ratelimit-requests-reset: 2025-07-23T15:05:09Z
anthropic-ratelimit-tokens-limit: 14000
anthropic-ratelimit-tokens-remaining: 12000
anthropic-ratelimit-tokens-reset: 2025-07-23T15:04:59Z

And for AWS Bedrock:

x-amzn-bedrock-invocation-latency: 2270

---

I can see how to access these in the Python SDK by using the with_raw_response e.g.:

>>> response = client.messages.with_raw_response.create(
... model="claude-3-5-sonnet-20241022",
... max_tokens=1000,
... messages=[{"role": "user", "content": "Hello"}]
... )
>>> response
<APIResponse [200 OK] type=<class 'anthropic.types.message.Message'>>
>>> response.headers
Headers({'date': 'Fri, 25 Jul 2025 14:28:57 GMT', 'content-type': 'application/json', 'transfer-encoding': 'chunked', 'connection': 'keep-alive', 'content-encoding': 'gzip', 'anthropic-ratelimit-input-tokens-limit': '20000', 'anthropic-ratelimit-input-tokens-remaining': '20000', 'anthropic-ratelimit-input-tokens-reset': '2025-07-25T14:28:57Z', 'anthropic-ratelimit-output-tokens-limit': '4000', 'anthropic-ratelimit-output-tokens-remaining': '4000', 'anthropic-ratelimit-output-tokens-reset': '2025-07-25T14:28:57Z', 'anthropic-ratelimit-requests-limit': '5', 'anthropic-ratelimit-requests-remaining': '4', 'anthropic-ratelimit-requests-reset': '2025-07-25T14:28:55Z', 'anthropic-ratelimit-tokens-limit': '24000', 'anthropic-ratelimit-tokens-remaining': '24000', 'anthropic-ratelimit-tokens-reset': '2025-07-25T14:28:57Z', 'request-id': 'req_011CRTtstudugSKpBy5C6idJ', 'strict-transport-security': 'max-age=31536000; includeSubDomains; preload', 'anthropic-organization-id': 'f8bf172e-0b01-42f5-9955-e0c43e1cb1a4', 'via': '1.1 google', 'cf-cache-status': 'DYNAMIC', 'x-robots-tag': 'none', 'server': 'cloudflare', 'cf-ray': '964c586c4a7b6405-LHR'})

But there doesn't seem to be an equivalent for Ruby

[dtmeadows]

Thanks for the report @kevindew! This is definitely something we plan to support in the Ruby SDK but haven't gotten around to implementing just yet. We'll update this issue once we have more to share.

[ms-jpq]

hi @kevindew I added some changes that exposes normalized headers, as well as the HTTP status for streams!

:D

9bfbae1#diff-85b3bc77621a3b82dde0414bac259815c4aefbcd34674385741626589f52488aR35

[jackbot]

Hi @ms-jpq, and thanks for looking into this. I'm struggling to figure out how to access the headers here - wonder if I'm doing anything wrong.

require 'bundler/inline'

gemfile do
  source 'https://rubygems.org'
  gem 'anthropic', '1.10'
  gem 'dotenv'
end

require 'anthropic'
require 'dotenv/load'

anthropic = Anthropic::Client.new(
  api_key: ENV['ANTHROPIC_API_KEY']
)

stream = anthropic.messages.stream(
  max_tokens: 1024,
  messages: [{ role: 'user', content: 'Hello, Claude' }],
  model: :"claude-sonnet-4-5-20250929"
)

puts 'Headers:'
puts stream.headers.inspect

This outputs "Headers: nil".

Seems like the headers are being set on raw_stream but then when instantiating a Anthropic::Streaming::MessageStream object (here) the headers are lost due to the overloaded initialize method in Anthropic::Streaming::MessageStream? That's just a guess though, I'm not too familiar with the code.

Similarly, should the headers exposed in this commit be available to non-streaming responses, and requests made via Bedrock?

[ms-jpq]

hi @jackbot you are right, will fix this 👍.

cc @renfredxh

[kevindew]

@ms-jpq could we reopen this issue please?

In case it's helpful we looked into where the headers get lost for the client.messages.create as well (which is the one we use on our implementation.

The headers are known straight after the HTTP response:

anthropic-sdk-ruby/lib/anthropic/internal/transport/base_client.rb

Line 508 in 7ef9279

headers = Anthropic::Internal::Util.normalized_headers(response.each_header.to_h)

But then are lost in the process of decoding the content:

anthropic-sdk-ruby/lib/anthropic/internal/transport/base_client.rb

Line 509 in 7ef9279

decoded = Anthropic::Internal::Util.decode_content(headers, stream: stream)

and not passed on further when the response model is created:

anthropic-sdk-ruby/lib/anthropic/internal/transport/base_client.rb

Lines 524 to 525 in 7ef9279

	unwrapped = Anthropic::Internal::Util.dig(decoded, unwrap)
	Anthropic::Internal::Type::Converter.coerce(model, unwrapped)

It would be helpful if the response models had a hash of headers as well as the stream object

[ms-jpq]

@kevindew yub, my apologies

[ms-jpq]

not gonna close the issue until 71fe5d0 is merged 👍

it is slated for next release 🙏

[kevindew]

Great to see there is progress.

Unfortunately as I understand that doesn’t satisfy this issue though as I believe that only provides header access for streams and not for non-streaming requests.

[kevindew]

@ms-jpq can we expect anything to change on this issue?

[ms-jpq]

hi @kevindew, I don't think we will have this before EoY, i am afraid :(