What actually changes when OpenAI publishes a new model_version?

When OpenAI ships a new dated snapshot (e.g., gpt-4o-2024-08-06 to gpt-4o-2024-11-20), the change can span any of seven layers. (1) Base weights — new pretrain or continued pretrain, the headline change. (2) Post-training — SFT + RLHF + DPO data updated, often where capability/refusal/style shifts originate. (3) System prompt defaults — OpenAI sometimes adjusts the implicit system prompt baked into the model, affecting tone, refusal calibration, and tool-use phrasing. (4) Tool-use schema — JSON-mode strictness, function-calling parameter validation, and structured-output adherence often tighten or loosen between snapshots. (5) Tokenizer or context window — rarely but consequentially, the tokenizer or effective context length changes. (6) Sampler defaults — implicit temperature/top-p priors at the serving layer can shift. (7) Safety classifier — input/output filters that decide refusal vs response are sometimes retrained.

OpenAI's release notes typically document (1), (2), capability deltas on internal benchmarks, and pricing changes. They rarely document (3)–(7) explicitly. Concrete recent examples: gpt-4o-2024-08-06 introduced strict structured-output mode (a (4)-level change); gpt-4o-2024-11-20 improved creative writing and instruction-following (a (2)-level change with measurable IFEval gains). To detect what changed in *your* workload, run a fixed eval suite against the old and new snapshot side-by-side — ModelWatch keeps both snapshots live in its rotation and surfaces a snapshot-diff view.