fix(openai): support o-series models in get_num_tokens_from_messages

[Anay Garodia (AnayGarodia)]

Closes #38700

---

Calling get_num_tokens_from_messages on a ChatOpenAI instance configured with any o-series model (o1, o1-preview, o1-mini, o3, o3-mini, o4-mini) raises NotImplementedError, because the model-prefix check that selects the per-message token format only recognizes gpt-3.5-turbo, gpt-4, and gpt-5 prefixes. Anyone counting tokens for budgeting or context-window management with reasoning models hits this immediately.

o-series chat models use the same message token accounting as the other post-gpt-3.5-turbo-0301 chat models (3 tokens per message, 1 per name), and tiktoken already resolves their encodings, so the fix is to include the o1/o3/o4 prefixes in the existing check.

The new parametrized unit test asserts each o-series model produces the same count as the reference chat model for identical messages and fails with NotImplementedError without the fix.

Release note

get_num_tokens_from_messages on ChatOpenAI now supports o-series reasoning models (o1, o3, o4 families) instead of raising NotImplementedError.

---

Disclaimer: this contribution was prepared with an AI coding agent (Claude Code), driven and reviewed by Anay Garodia (@AnayGarodia).

[Copilot]

Copilot was unable to review this pull request because the user who requested the review has reached their quota limit.