Anthropic model-info lookup does not resolve AWS Bedrock / cross-region inference model IDs #7833

Activity

1. Functionhx commented on Jul 9, 2026

   

   Yuchen Fan (Functionhx)

   on Jul 9, 2026

   More actions

   Added _normalize_model_id to strip Bedrock prefixes and version suffixes before model info lookup. Also filled in missing Claude 4 token limits. PR ready.

   👍React with 👍1Reacted by Lucas Kim
2. MOHAMMED WASIM KHAN (wasim-builds)

   added a commit that references this issue on Sep 2, 2026

   fix: resolve Bedrock-style model IDs in Anthropic model-info lookup

   ...

   This issue will close once commit 07ea995 is merged into the 'main' branch.

   07ea995
3. kimnamu commented on Sep 30, 2026

   

   Lucas Kim (kimnamu)

   on Sep 30, 2026

   Author

   More actions

   Thanks Eric Zhu (@ekzhu) for all the time you've put into autogen's model clients, including merging the Anthropic-on-Bedrock work in #6170 and #6979. This issue now has four open PRs for the same lookup fix (one of them is mine), so I've put them in one table to make it easier to pick one.

   PR	Author	New tests (file result)	Base	Workflow runs on head	Also in the PR
   #7834 (mine)	kimnamu	5 (23 passed)	main @ 027ecf0	0	none
   #7886	C1-BA-B1-F3	7 in a new test file (25 passed)	main @ 027ecf0	0	none
   #7930	HumphreySun98	3 (21 passed)	main @ 027ecf0	7, waiting for approval	adds the missing Claude 4 entries to _MODEL_TOKEN_LIMITS
   #8144	wasim-builds	4 (22 passed)	main @ 027ecf0	7, waiting for approval	also targets #7867 (tool output serialization) and #7768 (trailing assistant messages)

   All four sit on current main with no conflicts, and all four fix the reported KeyError. In my run the visible difference was the Claude 4 token limit, where only #7930 returns 200000:

   get_info / get_token_limit("us.anthropic.claude-sonnet-4-20250514-v1:0")
   main @ 027ecf0   KeyError          100000
   #7834            claude-4-sonnet   100000
   #7886            claude-4-sonnet   100000
   #7930            claude-4-sonnet   200000
   #8144            claude-4-sonnet   100000
   

   ErenAta16 also compared #7834, #7886 and #7930 head to head in a review on #7930, including a us-gov. prefix case I haven't repeated here.

   When you have a moment, could you say which one you'd like to take forward? If this isn't a priority right now, or you'd rather go with one of the others, tell me and I'll close #7834 to keep your queue clean.

   How this was run

   Each PR's _model_info.py was loaded from its head commit and called with the same IDs. Test results are tests/models/test_anthropic_model_client.py (plus test_anthropic_model_info.py for #7886) in python/packages/autogen-ext, run without API credentials: 18 passed on main, 20 skipped on every branch.

   id                                               main                       #7834                      #7886                      #7930                      #8144
   anthropic.claude-3-5-sonnet-20240620-v1:0        KeyError, 100000           claude-3-5-sonnet, 200000  claude-3-5-sonnet, 200000  claude-3-5-sonnet, 200000  claude-3-5-sonnet, 200000
   global.anthropic.claude-sonnet-4-20250514-v1:0   KeyError, 100000           claude-4-sonnet, 100000    claude-4-sonnet, 100000    claude-4-sonnet, 200000    claude-4-sonnet, 100000
   claude-3-5-sonnet-20240620                       claude-3-5-sonnet, 200000  claude-3-5-sonnet, 200000  claude-3-5-sonnet, 200000  claude-3-5-sonnet, 200000  claude-3-5-sonnet, 200000