Speech and audio models are splitting along price. The general-purpose end is collapsing toward commodity rates while specialized work commands a premium.
Anthropic released Claude Haiku 5.5 and halved cache-read prices on Sonnet 5.5, continuing a pattern of pushing capable small models into workloads that do not need top-tier reasoning. OpenAI has been advertising enterprise wins on the same economics, including Sophos cutting threat investigation time by 96 percent and Asana reporting 76-fold model cost reductions on browser tests.
That tier now handles transcription, summarization and routine voice interaction at prices that make long-running pipelines viable.
AudioShake is building at the other end, launching The Refinery to convert overlapping conversations into usable training data. The product targets recordings where several people talk at once and a single transcript is worthless, separating and labelling speakers so the output can feed downstream models.
Real-world audio is where general models struggle most. Cafes, call centers and field recordings carry crosstalk, noise and clipped speech that clean benchmarks do not represent.
The split has a clear logic. Cheap models absorb volume, and anything requiring source separation or forensic accuracy gets handled by tooling built for it.
