weiqingy commented on PR #1039:
URL: https://github.com/apache/flink-agents/pull/1039#issuecomment-5467168270

   Thanks, all three are fixed in cb4453ab.
   
   The default is now `claude-sonnet-4-6`. Your point about the alias not 
advancing past 4.5 is right, and the docs back it a step further: 
`claude-sonnet-4-6` is listed as the recommended replacement for 
`claude-sonnet-4-20250514`, which is the exact value this default used to carry.
   
   The PR first picked 4.5 because of the JSON prefill, which 4.6 rejects. That 
turned out not to matter. Both runtimes already gate the prefill on the model 
(`AnthropicChatModelConnection.java:425`, `anthropic_chat_model.py:376`), so it 
is suppressed rather than sent. 4.6 also matches the native structured-output 
allowlist exactly, where the 4.5 alias only matched the prefix rule, so callers 
on the default now get the native path instead of the prompt-engineering 
fallback.
   
   I stopped at 4.6 rather than going to the 5 line because `temperature` is 
always sent, and 4.7 and later reject a non-default sampling value with a 400. 
#1037 is adding that gating and would lift the ceiling. Does moving to 4.6 now 
and revisiting after #1037 seem like the right order, or would you rather hold 
and move once?
   
   On requiring an explicit model instead, I lean toward keeping a default, 
since every other chat model integration here ships one (`gpt-4o-mini`, 
`gpt-4o`, `gemini-3.1-pro-preview`, `qwen-plus`). Is there something about 
Anthropic that you think warrants treating it differently?
   
   You were right about the Python test. It compared against the constant it 
imported, so it checked the wiring but not the shipped value. It asserts the 
literal now, matching the Java test. `ChatModelIntegrationAgent` is fixed too, 
and it was worse than stale: `claude-sonnet-4-20250514` retired on June 15, 
2026, so that path could not have run against a live key at all. I also 
refreshed the docs list of popular models, which named only 4.5-generation 
entries.
   


-- 
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.

To unsubscribe, e-mail: [email protected]

For queries about this service, please contact Infrastructure at:
[email protected]

Reply via email to