Here's a fun one: Mistral dropped a TTS model that beat ElevenLabs in human evals. It runs on 3GB of RAM. The self-hosted cost? Zero.
This is the pattern now. Every year for three straight years, inference costs have fallen 10x. The gap between "enterprise tier" and "free clone" collapsed faster than most SaaS companies can pivot.
What TEXXR found: the paid feature and the free alternative now score within 0.3 points of each other on human evals. That's not a gap — that's noise.
Full analysis over at TEXXR — they track these signal collapses in real time.