The cost of using frontier AI models has quietly collapsed. Since June, OpenAI and Anthropic have cut the prices of their flagship LLMs by 40% to 80% through a rapid series of overlapping price reductions — and an analysis published October 25 by data scientist Max Woolf lays out, month by month, how a full-blown price war unfolded between the two companies.
The timeline matters because it shows the cuts were not a one-time discount but a sustained competitive spiral, with each company responding to the other's moves within days. For more context on this story, see our ongoing AI news coverage.
The Timeline: Four Months of Falling Prices
Woolf's analysis tracks per-million-token pricing — input, output, and cache read — across every major release from both labs since late June:
- June 26: OpenAI launched GPT-5.6 Sol at $5 input / $30 output per million tokens, with small model Luna at $1 / $6.
- June 30: Anthropic launched Claude Sonnet 5 at $2 / $10 — a launch price it described as temporary.
- July 24: Anthropic launched Claude Opus 5 at $5 / $25, directly matching GPT-5.6 Sol.
- July 30: OpenAI cut GPT-5.6 Luna's price by 80%, to $0.20 / $1.20.
- August 10: Anthropic made Sonnet 5's discounted price permanent.
- August 21: OpenAI cut GPT-5.6 Sol by roughly 20%, to $4 / $20, for three months.
- September 1: Anthropic launched Fable 5.1 at $10 / $50 with the cache-read price cut by 75%.
- September 22: OpenAI launched GPT-6 Sol at a startling $2 / $10 — half its discounted predecessor — and on the same day, Anthropic launched Claude Opus 5.5 at $4 / $20, with the company stating that "Opus 5.5 requires less compute to serve than Opus 5, and its pricing reflects that."
- September 29: OpenAI followed with GPT-6.1 Sol, again at $2 / $10 but with the cache-read price halved to $0.10 — making the model family cumulatively about 75% cheaper than in June.
By Woolf's accounting, the cost of using these model families has cratered in four months, with no visible sacrifice in output quality. The cuts also flow through subscription-based coding tools like Codex and Claude Code, not just raw API access.
Cache Pricing Is the Hidden Battleground
One of the analysis's sharpest insights concerns cache read costs — the price of re-reading a model's memory of an ongoing conversation. In agentic coding workloads, cache reads make up 96% to 99% of total token expenditure, and cache hit rates typically run 90% to 100%.
That arithmetic is why Anthropic's 75% cache-read cut for Fable 5.1 correlated with a 70% to 75% reduction in the real cost of using the model, even though the headline input and output prices never changed. When OpenAI halved GPT-6.1 Sol's cache price a month later, it was matching Anthropic's cache set at roughly 5% of the input price.
Why Cut Prices 80%?
Woolf's answer centers on commoditization. Since agentic LLM usage entered hypergrowth in November 2025, he argues, models have become easy to swap: developers can switch between models with a toggle in tools like OpenCode or a different endpoint on OpenRouter, which leaves labs with little moat and plenty of pricing pressure.
The other externality is open-weight competition. DeepSeek's V4 Flash, released in April 2026 at $0.14 input / $0.28 output with a cache-read price of just $0.0028, went viral precisely because OpenAI lacked a cheap small-class model at the time. Woolf suspects the 80% GPT-5.6 Luna cut in July was aimed squarely at these smaller open-weight models. DeepSeek's updated Flash V4.1 scores 45 on the Artificial Analysis index, up from 24, at similar prices — keeping the pressure on.
"No one cuts prices by 80% unless there is some externality in play," Woolf writes.
Spending Is Rising Anyway
The twist at the end of the analysis: cheaper tokens have not meant cheaper bills. Data from Vercel's AI model gateway shows the cost per token falling to 43.3% of its April 2026 level — while aggregate spending keeps climbing, because models are getting more capable faster than they are getting cheaper.
That dynamic explains the strategic logic on both sides. For the labs, lower prices grow usage and lock in developers even as margins per token compress. For customers, the effective price of a unit of useful work — a closed ticket, a merged pull request, a completed research task — is falling even faster than the sticker prices suggest.
What to Watch Next
If the pattern holds, the next model releases from OpenAI and Anthropic will arrive with headline prices at or below their predecessors, and cache-read pricing will keep converging toward the aggressive ratios DeepSeek established. Google and Meta, which Woolf notes also compete in this market, face the same math.
For now, buyers of AI compute are the clear winners of a war that shows no sign of ending — and the analysis suggests the endgame is a market where frontier capability is priced like a commodity, with differentiation coming from everything wrapped around the model rather than the model itself.
---
Stay Ahead of AIGet the latest AI news, analysis, and breakthroughs — all in one place.
Read more AI news →
