Flagship LLM prices fell up to 80% — cache reads are your real bill
A fresh breakdown of 2026's price war: what each flagship costs per token now, why the cuts keep coming, and the open-weight leverage that drives them.
Copy markdown
Cache reads are 96–99% of an agent's token bill
In Max Woolf's accounting, cache reads — not input or output — make up 96–99% of coding-agent token spend, so the rate to optimize is the cache-read price and your hit rate, not the headline input number. GPT-6.1 Sol just halved its cache read to $0.10/MTok and Claude Opus 5.5 sits at $0.20; structure prompts so the cache actually hits before you chase a cheaper input tier.
What each flagship costs per million tokens now
GPT-5.6 Luna fell 80% to $0.20/$1.20/$0.02 (in/out/cache); GPT-6 Luna runs $0.10/$0.50/$0.01; Claude Sonnet 5 is $2/$10/$0.20 and Opus 5.5 is $4/$20/$0.20; DeepSeek V4.1 Flash undercuts them all at $0.15/$0.60/$0.003. Re-price your pipelines against today's rates instead of spring numbers — the cheap tier is now genuinely cheap.
Open weights are why the bill keeps dropping
With ~29 providers serving a single DeepSeek model on OpenRouter and near-zero switching cost, hosted vendors are locked in a price-cut tit-for-tat; Vercel's data has cost-per-token at 43% of its April 2026 level. Benchmark an open-weight alternative before you renew — that leverage is doing the work.
Cheaper tokens won't shrink your bill — watch your plan limits
The caution worth pinning: models are improving faster than prices fall, so usage and total spend tend to climb even as per-token rates drop — and subscription allowances have quietly been trimmed alongside some cuts. Budget for consumption growth, and re-read your plan's quota fine print before you bank the savings.