OpenAI and Anthropic cut LLM prices by up to 80% in four months, and cache-read cuts explain most of the drop
- On July 30, 2026, OpenAI cut GPT-5.6 Luna by 80% to $0.20 / $1.20 / $0.02 per MTok input, output and cache read, days after Anthropic made Sonnet 5's launch pricing permanent.
- Cache read pricing drove most of the real savings: Anthropic's Fable 5.1 cut cache reads from $1 to $0.25 per MTok, and since cache reads are 96% to 99% of token spend in agentic coding, that took 70% to 75% off total usage cost.
- Claude Opus 5.5 (September 22, 2026) at $4 / $20 / $0.20 cut total cost 40% to 60% against Opus 5, and Anthropic stated outright that "Opus 5.5 requires less compute to serve".
- GPT-6.1 Sol (September 29, 2026) halved its cache price to $2 / $10 / $0.10 to match Anthropic's 5%-of-input cache rate, making the GPT-6 family cumulatively 75% cheaper than where it started in June.
- The author cites Vercel's AI model gateway data showing cost per token down to 43.3% of the April 2026 level while aggregate spend keeps rising, and argues output quality was not traded away for the cuts.
Hacker News 의견들
I fully expect token vendors to end up like mobile data providers: throughput keeps climbing, everything commoditizes, and they never capture the gains their own tech creates. Can't wait for the unlimited plans.
Unlimited plans won't come. You can only talk on the phone so long, but hand me unlimited tokens with scheduled jobs and subagents and I'll burn a fortune.
Serving a model scales linearly with use because of energy. Serving a mobile data plan costs roughly zero, so the flatrate comparison falls apart.
Competition drives this. We kept hitting cybersecurity refusals on OpenAI and Anthropic, switched to Kimi and GLM, and found them much cheaper while finishing the work without refusals.
If it's mostly competition, either their usage has cratered or they implode soon. They were already burning money at higher prices and capex wouldn't have shrunk proportionally.
Imagine a beef shortage and the beef producers cutting prices because of competition. Something doesn't add up here.
Price war, obviously. There's no moat, and the Chinese labs are pushing too. It's a race to the bottom, ugly for the stock market and great for consumers.
Great for consumers only if you want more stuff rather than better stuff. The bubble popping hurts everyone, and so far all this has done is make everything digital feel less trustworthy and less safe.
It's probably better caching plus smaller models. GPT-6 Sol is really 6 Terra and Opus 5.5 is smaller than 5.1. Don't overthink it.
GPT-6 Sol was a dud, and a week later 6.1 Sol showed up. I'm convinced Sol was only ever meant to be the web chat model and got panic released to answer Opus 5.5.
People act like OpenAI and Anthropic have no idea what they're doing. The naysayers sound exactly like the ones who said this internet thing wasn't going anywhere.
They aren't making profit right now, that's just a fact. And nobody serious claims LLMs are useless, the question is whether a viable business model exists.
Global Crossing built the internet backbone and went bankrupt. The internet survived, most of the giant companies didn't.
This is exactly why they're pushing for more regulation. You can't show investors ROI by cutting inference costs, but you can if Chinese open models get banned.
A lot of it, maybe most of it, is KV cache optimization.
That link is just some random blogger's overconfident theory.