minimaxir: the price cuts are about cache reads, and they won't reverse
Max Woolf, who blogs as minimaxir, put every OpenAI and Anthropic price change since June 26 on one dated page and asked the obvious question: why did both labs cut up to 80% within a few months? The answer in the post is boring and important. For coding agents, the bill is mostly cache reads.
The number doing the work: "When using agents for coding, cache read costs are the vast majority (96% - 99%) of the total token expenditure", because the cache holds the whole growing context and gets re-read every turn. So when Anthropic launched Fable 5.1 at $10 input, $50 output, $0.25 cache read and cut the cache read price by 75%, the effective cost of using the model fell about 70 to 75%, even though the headline input and output prices look expensive. Opus 5.5 at $4 / $20 / $0.20 works out to a 40 to 60% total cut by the same math. GPT-6 Sol sits at $2 / $10 / $0.20.
The second half is the commodity argument. DeepSeek V4.1 Flash runs at $0.15 / $0.60 / $0.003 across 29 providers on OpenRouter, up from 13 in May. Kimi K3.1 Pro is priced at $2 / $10 / $0.20 with an Artificial Analysis score of 55 against Opus 5.5's 58, and its weights are due November 4, after which most of the 20-plus Kimi K3 hosts can serve it too. The post's conclusion: "The idea that token prices will suddenly rise is more a fantasy." The moats it names are enterprise commitments and personal agents like OpenAI Dots and Claude Duet, not tokens.
The useful takeaway for builders is which number to watch. Comparing models on input and output price per million tokens is now close to meaningless for agent workloads. The real cost is cache-read price times your cache hit rate times how long your context gets, which is a harness design question as much as a vendor question. A harness that compacts well or keeps prefixes stable can save more than switching labs.
Post: https://minimaxir.com/2026/10/cheap-llms/ HN: https://news.ycombinator.com/item?id=50236102
← Back to all articles
The number doing the work: "When using agents for coding, cache read costs are the vast majority (96% - 99%) of the total token expenditure", because the cache holds the whole growing context and gets re-read every turn. So when Anthropic launched Fable 5.1 at $10 input, $50 output, $0.25 cache read and cut the cache read price by 75%, the effective cost of using the model fell about 70 to 75%, even though the headline input and output prices look expensive. Opus 5.5 at $4 / $20 / $0.20 works out to a 40 to 60% total cut by the same math. GPT-6 Sol sits at $2 / $10 / $0.20.
The second half is the commodity argument. DeepSeek V4.1 Flash runs at $0.15 / $0.60 / $0.003 across 29 providers on OpenRouter, up from 13 in May. Kimi K3.1 Pro is priced at $2 / $10 / $0.20 with an Artificial Analysis score of 55 against Opus 5.5's 58, and its weights are due November 4, after which most of the 20-plus Kimi K3 hosts can serve it too. The post's conclusion: "The idea that token prices will suddenly rise is more a fantasy." The moats it names are enterprise commitments and personal agents like OpenAI Dots and Claude Duet, not tokens.
The useful takeaway for builders is which number to watch. Comparing models on input and output price per million tokens is now close to meaningless for agent workloads. The real cost is cache-read price times your cache hit rate times how long your context gets, which is a harness design question as much as a vendor question. A harness that compacts well or keeps prefixes stable can save more than switching labs.
Post: https://minimaxir.com/2026/10/cheap-llms/ HN: https://news.ycombinator.com/item?id=50236102
Comments