Switch default parser and summarizer from gpt-4.1-mini to Claude Haiku 5.5

[CodexVeritas]

Summary

OpenAI retires gpt-4o, gpt-4o-mini, gpt-4.1 and gpt-4.1-mini on Jan 14th 2027. Two shared defaults that every bot in run_bots.py runs through use gpt-4.1-mini:

• DEFAULT_STRUCTURE_OUTPUT_MODEL (the parser for all bots, and the library default for structure_output)
• default_summarizer in run_bots.create_bot (summaries are shown in comments only; use_research_summary_to_forecast is False)

Both now use openrouter/anthropic/claude-haiku-5.5. Dated history comments were added in run_bots.py so bot results can be traced to the parser/summarizer in use.

Model comparison

structure_output with num_validation_samples=2 (same as the template bots), 5 reps per case. 11 cases: the 7 from test_structure_output.py plus 4 bot-style cases (binary with a revised draft probability, binary with a decimal percent, multiple choice with Option_X prefixes and a 0% option, numeric needing $ → M $ conversion and scientific notation).

Model	Passed	Cost (55 runs)	Median latency	Retries
openai/gpt-4.1-mini (old)	55/55	$0.077	2.0s	0
anthropic/claude-haiku-5.5	55/55	$0.041	3.1s	5
openai/gpt-6-luna	53–54/55	$0.023	4.5s	7
google/gemini-3.1-flash-lite	50/55	$0.068	1.7s	0

All Haiku and Luna retries came from one case (bare float output with "give your answer as a percentage", where the model answers 30%). Bots parse into pydantic types, where Haiku had no retries. Luna's failures were also on that case; Gemini returned 0.3 instead of 30.

Testing

• code_tests/integration_tests/test_structure_output.py passes with the new default
• Unit tests pass

Bot definitions that use the retiring models as their own forecasting LLMs are unchanged.

🤖 Generated with Claude Code