[aksg87]

Description

Summary

• The OpenAI provider's default model_id, gpt-4o-mini, resolves to
  gpt-4o-mini-2024-07-18, which OpenAI shuts down on January 14, 2027
  (recommended replacement: gpt-6-luna). After that date every call that
  relies on the default fails.
• Changes the default to gpt-6-luna and updates the README examples,
  docstrings, the provider README and the bundled usage skill docs, including
  a note that temperature needs reasoning_effort="none" on this model.
  Also replaces two examples that named already-retired models (gpt-4 in
  langextract/providers/README.md, gpt-3.5-turbo in the skill docs).
• Depends on Route gpt-6 model IDs to the OpenAI provider #599 (the README and skill examples call
  lx.extract(model_id="gpt-6-luna")) and Send max_output_tokens as max_completion_tokens to OpenAI #600 (otherwise callers who set
  max_output_tokens with the new default would send max_tokens). Draft
  until both merge and the release that ships it is decided.

Feature

Compatibility

• This only affects callers who omit model_id, for example
  OpenAILanguageModel(api_key=...) or
  ModelConfig(provider="OpenAILanguageModel"). Callers who pass model_id
  see no change. To keep the old behavior until the shutdown, pass
  model_id="gpt-4o-mini".
• gpt-6-luna is a reasoning model with medium effort by default
  (model page),
  so latency and output tokens change (reasoning tokens are billed as
  output). List price is $0.10 input / $0.50 output per 1M tokens, versus
  $0.15 / $0.60 for gpt-4o-mini. Pass reasoning_effort="none" or "low"
  to stay closer to the old behavior.
• Per OpenAI's
  GPT-6 guide,
  temperature, top_p, top_logprobs and logprobs are accepted only with
  reasoning_effort="none". Callers who set temperature and rely on the
  default must also pass reasoning_effort="none".
• Reasoning tokens count against max_output_tokens, so a small budget with
  the default model can now end with finish_reason=length and no text.
• In Batch API mode, rejected parameters surface as per-request errors after
  the batch completes rather than immediately.
• OpenAI-compatible servers (base_url) that relied on the default model name
  must pass model_id explicitly.

How Has This Been Tested?

Tests

• pytest -m 'not live_api and not requires_pip and not integration':
  983 passed, 25 deselected.
• Updated default assertions: test_openai_temperature_zero in
  tests/inference_test.py (now also passes reasoning_effort="none", the
  only form the API accepts with temperature) and
  test_explicit_provider_selection in tests/test_live_api.py.
• isort 5.13.2 and pyink 24.3.0 checks are clean; pylint is clean on the
  changed files.
• Not verified here: no live test sends a request through the default model;
  test_explicit_provider_selection only constructs it. Move OpenAI live tests off retiring GPT-4o models #601 covers
  gpt-6-luna requests, including the output-schema test at default effort.

Checklist

• I have made any needed documentation changes.
• I have added tests, or I have ensured existing tests cover the changes.
• I have followed Google's Python Style Guide and ran pylint over the
  affected code.