AI inference pricing is entering a race to zero
Meta's Muse Spark 1.1 launched with pricing at $1.25/$4.50 per million tokens (input/output) and cached reads at just $0.15, which the HN thread called 'insane' by current market standards. Commenters immediately stacked it against Grok 4.5 and found it cheaper. The thread is framing this as a deliberate competitive move, not just a product launch.
The pattern across multiple AI threads today: GPT-5.6 is being mocked as a decimal-point update to stay in the news cycle, Hy3 is described as shockingly capable for its size, and GLM 5.2 is being run on slow hardware with swap files. The market is bifurcating. At the top, labs are competing on price. At the bottom, smaller models are getting good enough to run locally on consumer hardware. Both trends squeeze the middle tier of API-dependent startups who built on premium model pricing.
The counterpoint from the community: more competition is genuinely good for builders. The list getting cited is Chinese models, Grok, Meta, Google, OpenAI, Anthropic. Six serious competitors is a real market, not a duopoly.
So what?
If you're building a product on top of a specific model API, your unit economics are about to change, probably downward on cost but upward on competitive pressure. The smarter founders are watching Meta's open-source strategy specifically: cheap API plus open weights means you can switch or self-host, which is real negotiating leverage against OpenAI and Anthropic. Lock-in is getting harder to enforce.