LGTM!

What does this merge request do and why?

Now that long-context window is GA for Sonnet 4.6 and Opus 4.6 (see https://claude.com/blog/1m-context-ga), context-1m-2025-08-07 beta headers are no longer needed for Sonnet 4.6 and Opus 4.6

From Anthropic:

The 1M token context window is now generally available for Claude Opus 4.6 and Sonnet 4.6 via API. Both models include the full window at standard pricing—$5/$25 per million tokens for Opus 4.6 and $3/$15 for Sonnet 4.6. Previously there were separate rate limits above and below 200K tokens. We’ve simplified this to a single rate limit for the full context window. As part of this, we’ve raised your base rate limit on Opus 4.6 to 18M to accommodate your existing long context usage with room to grow. Note this applies only to the Gitlab Production (managed account) org. As always, please reach out to request additional increases.

We are also updating pricing multipliers in https://gitlab.com/gitlab-org/customers-gitlab-com/-/merge_requests/15070

Numbered steps to set up and validate the change are strongly suggested.

Merge request checklist

• Tests added for new functionality. If not, please raise an issue to follow up.
• Documentation added/updated, if needed.
• If this change requires executor implementation: verified that issues/MRs exist for both Go executor and Node executor or confirmed that changes are backward-compatible and don't break existing executor functionality.

@wortschi fair point. Should be fine if we already have context info in the MR description.

@romaneisner would you be able to help review it?

Add two comments.

suggestion (non-blocking): it would be great to add anthropic announcement or links to raise awareness that this feature is exclusively support to Opus and Sonnet 4.6 only.

suggestion (nit-picking): 1_000_000 format generally has a better readability, I don't need to count the number of 0 😀.

Hi @bastirehm @wortschi @timzallmann , now the Claude 1m context window GA at a standard pricing, we need to add support on ai-gateway to leverage it. cc @achueshev

Problem to solve

As a GitLab Duo user, I want to leverage the full 1M context window now available for Claude Opus 4.6 and Sonnet 4.6, so I can work with larger codebases and documents in a single conversation without context truncation.

Anthropic announced general availability of 1M context window (March 13, 2026) with:

• Standard pricing across full 1M window (no long-context premium)
• Full rate limits at every context length
• No beta header required for requests over 200K tokens

Proposal

Update AI Gateway to support the full 1M context window for Claude Opus 4.6 and Sonnet 4.6 models.

Implementation Plan

1. Update Model Configuration

   • Update models.yml to reflect 1M (1,000,000) token context window for Claude Opus 4.6 and Sonnet 4.6
   • Remove any 200K context limit restrictions

2. Remove Beta Header Requirements

   • Identify and remove any anthropic-beta header logic for long-context requests
   • Requests over 200K tokens should work automatically

3. Testing

   • Add/update tests for 1M context window support
   • Verify no regression in existing functionality

Further details

Benefits:

• Users can process entire codebases in a single session
• Improved accuracy for large document analysis (Opus 4.6 scores 78.3% on MRCR v2)
• No additional cost for long-context requests

Acceptance Criteria:

• Claude Opus 4.6 and Sonnet 4.6 accept up to 1M tokens in context
• No beta headers required for long-context requests
• All existing tests pass

Links / references

• Anthropic announcement: https://claude.com/blog/1m-context-ga

Hey @bluenoodles, do you have any progress or hit any blockers on this? This feature turns out needed to be prioritized soon. If you are unable to deliver, feel free to let us know. We could also suggest other issues that you can work on.

@bcardoso- would you be able to do the maintainer review of this short adding compaction doc MR?

Addressed

Junming Huang (4bb1219b) at 16 Mar 13:22

@GitLabDuo that should be addressed by looking at the yaml example and code itself.

@GitLabDuo should be fine.

Hi @alejandro @fpiva, we are adding conversation history compaction feature to DAP which requires making llm call that is not triggered by user. I have the following questions hope you can help:

1. Are we already able to filter llm request when doing usage billing (bill customer / not bill)? If yes, how can we usage such feature? If not, do you know if it is something already on the usage billing roadmap and how complex is that?
2. Currently, the compaction llm call is using model from the workflow with hard coded system prompt template. I would like to migrate to use prompt registry, does it support to get the raw llm model object (without tool binding)? Is there any particular things I need to pay attention to?

cc @bastirehm @wortschi

fixed

Junming Huang (790ef606) at 16 Mar 13:04

Junming Huang (9f07b662) at 16 Mar 12:50

@eduardobonet can you help do the initial review of this small MR?