The Data Behind the LLM Price War: OpenAI and Anthropic Cut Costs Up to 80% in Three Months

Max Woolf’s analysis, published October 25 and front-paged all morning, puts hard numbers under the discount war we covered yesterday: tracing every major model announcement from June 26 through late September plus pricing data from Vercel’s AI gateway and OpenRouter, the conclusion is that the GPT-6.1 Sol family ended up roughly 75% cheaper cumulatively, Anthropic’s Sonnet 5 launch discount of $2/$10 became permanent in August, and across the board the price of frontier intelligence fell by up to 80% in a single quarter.

The cache insight is the real story

The most valuable part of the analysis is an economics lesson most price announcements bury: for agentic coding workloads, cache reads are 96 to 99% of token spend, so Anthropic’s September 1 cut of Fable 5.1’s cache-read price by 75% (to $0.25 per million) yields roughly a 70 to 75% total cost reduction even with headline rates unchanged. OpenAI’s sequence tells the same story: Luna launched June 26 at $1/$6, was slashed 80% to $0.20/$1.20 on July 30 in what Woolf links to DeepSeek’s cheap-cache pressure, and by September 22 the GPT-6 generation arrived at $0.10/$0.50 with a cache price of a cent. Whoever prices the cache wins the agent market, because the agent is the workload.

Four causes, one dismissed

Woolf attributes the cuts to four forces: tit-for-tat competition with a visibly reactive announcement pattern between OpenAI and Anthropic; open-weight pressure, with DeepSeek’s cache costs at roughly 2% of input price and 29 competing providers on OpenRouter routing toward the cheapest option; genuine efficiency gains, where Anthropic’s claim that Opus 5.5 simply costs less to serve is plausible; and possible IPO positioning, though slashing your main revenue line before an S-1 seems counterproductive, and per-token inference margins may remain near 80% regardless. He then dismantles the loss-leader thesis directly: unlike the Uber-Lyft duopoly, no pricing power exists here, open models are hosted at cost by dozens of providers, switching costs are negligible, and no provider hosts unprofitable models voluntarily, so prices should trend toward marginal cost and stay there.

What it means for anyone building on LLMs

Three practical conclusions follow from the data. First, aggregate spend is still rising: Vercel’s gateway shows costs at 43.3% of April 2026 levels while total spend climbs, so models are getting cheaper faster than they are getting used less. Second, quality was not sacrificed; the cheap models are the new generations, not downgraded variants, which is why Rippling’s workload could flip on price alone. Third, watch the subscription tier instead of the API: OpenAI halved the $200 Codex plan’s allowance the same month it cut API prices, meaning the value extraction is migrating to where users have no per-token visibility. Woolf’s closing point is the strategic one: even if both companies stumbled at IPO, the open-weight ecosystem and its at-cost hosts would keep the market honest, which is the opposite of the lock-up the bank-blocking and export-control stories are writing around the edges.