An independent analyst argues that openly published models have become good enough for most agent-style coding work, and that available computing capacity rather than model quality now sets prices. OpenAI has cut prices on both its cheap and flagship tiers, while Meta offers its open model at near-zero rates on a tier where it may train on submitted data. Anthropic's most expensive model is reported to be struggling to attract users at $10/$50 per million tokens.
What changed
Frontier proprietary models commanded a clear quality premium and set pricing.
What it unlocks
Running agent-style work on openly available models at roughly a tenth of frontier prices.
- GPT-5.6 Luna cut 80%
- Sol cut 20% for three months
- Muse Spark 1.2 at $0.10/$0.20 per MTok
- Fable 5 at $10/$50 per MTok
What you need to act on it
- a hosting provider or contributor tier where the vendor may train on submitted data
Sources