What's this about?
Three pricing moves surfaced on the same day. Grok 4.6 quietly raised its cache-read price to match GPT-5.6 Sol, days after launching on a "pricing unchanged" message. DeepSeek V4 Pro is charging as much as 12 times more during peak hours. And Gemini 3.7 Flash's headline price is explicitly introductory, scheduled to double on December 31, 2026. Simon Willison called that last one genuinely weird: who plans to still be using this model in five months, when its predecessor was only three weeks old?
What exactly is "cache read" pricing?
When you use a coding agent, most of what you send is identical every turn — the same files, the same system instructions. Providers store that processed context and charge a much lower rate to re-read it than to process it fresh. This matters more than it sounds: in long coding sessions, cache reads can be around 80% of the total bill. So a provider can leave the headline input and output prices untouched, raise only the cache rate, and still substantially increase what heavy users actually pay — which is precisely what a Grok user documented on HN.
What's the community saying?
- The headline price is no longer the price. The connecting thread across all three: the number in the announcement and the number on the invoice are diverging, via cache rates, peak-hour surcharges, and expiry dates.
- "Elon's not playing fair." OrangeAI's reaction to the Grok cache increase — "you raised it to match GPT-5.6 Sol; do you think you're Sol?" — captures the objection that a price rise should follow a capability claim, not precede one.
- Peak-hour pricing is new territory. DeepSeek charging up to 12x at peak imports a utility-style pricing model into LLM APIs, which makes cost forecasting much harder for anyone building a product on top.
- Short model lifespans make forward pricing meaningless. Willison's point cuts deeper than the joke: when a tier is replaced every three weeks, a promise about December is a promise about a model nobody expects to be running.