Recommended Articles
Tech & AI

AI Inference Prices Fell Up to 80% in Four Months — and Founders Should Reprice Everything

By Angeled · October 26, 2026 · Tech & AI
AI Inference Prices Fell Up to 80% in Four Months — and Founders Should Reprice Everything
AI Inference Prices Fell Up to 80% in Four Months — and Founders Should Reprice Everything

Between June and September 2026, the cost of running leading AI models collapsed. OpenAI and Anthropic cut prices on their flagship models by up to 80%, according to an analysis by Max Woolf (minimaxir) published in October 2026 and widely discussed on October 26. The price war was driven by competition from open-weight models — above all DeepSeek — by rapid gains in inference efficiency, and by the simple economics of perishable GPU capacity. For founders building on AI, the cost structure of the last two years has been rewritten in a single quarter.

The timeline of the collapse

The cuts came in a rapid sequence. On June 26, 2026, OpenAI announced GPT-5.6 Sol and Luna; Luna launched at $1 per million input tokens and $6 per million output tokens. On June 30, Anthropic announced Claude Sonnet 5 at $2/$10 launch pricing, initially described as temporary. On July 24, Anthropic launched Claude Opus 5 at $5/$25. Then on July 30, OpenAI cut Luna's price by 80%, to $0.20/$1.20. Anthropic made Sonnet 5's launch pricing permanent on August 10, and on September 22 announced Claude Opus 5.5 at $4/$20 — a 20% input/output cut plus a much larger cache-read cut. OpenAI responded with GPT-6.1 Sol, halving cache-read prices and making the family cumulatively about 75% cheaper.

The single most important detail is cache reads. In agentic coding workflows — tools like Codex and Claude Code — cached conversation history is re-read every turn, and those reads account for 96–99% of total token spend. A 75% cut in cache-read pricing therefore translates into a 70–75% cut in effective cost for developers, even when headline input/output prices move less.

Why prices fell: DeepSeek, efficiency, and perishable capacity

Woolf's analysis points to three drivers. First, Chinese competition. DeepSeek's Flash V4 is reportedly priced around $0.60 per million output tokens, and Woolf notes that DeepSeek's Flash V4.1 scores 45 on the Artificial Analysis index versus 24 for Flash V4 at $0.15/$0.60. These are reported figures, not confirmed list prices, but the direction is clear: open-weight models served by dozens of inference providers on routing platforms like OpenRouter have set a price ceiling that frontier labs cannot ignore. Woolf notes that without the Luna cut, OpenAI would have lost to DeepSeek V4 Pro on cost-versus-quality measures.

Second, efficiency. Both labs improved inference performance — reduced compute per token in Opus 5.5, for example — allowing them to cut prices without sacrificing output quality. Third, capacity economics: GPU capacity is perishable like airline seats. An idle GPU earns nothing, so providers would rather sell tokens cheaply than leave hardware unused. Data from Vercel's AI model gateway shows cost per token falling to 43.3% of April 2026 levels, even as aggregate spending continues to rise because usage grows faster than prices fall.

What cheaper inference means for founders

The implications run in several directions. First, unit economics flip. Business models that were marginal at 2025 token prices — AI customer support, document processing, coding agents, voice agents — may now be profitable, and founders should re-run their cost models rather than assuming old prices. Second, differentiation moves up the stack. When the model itself is cheap, the value shifts to workflow, data, distribution, and trust — the things a customer cannot get by switching providers. Third, pricing strategy matters more than ever: if your product's price is anchored to token costs, your margins will compress as fast as your inputs cheapen unless you price on outcomes.

There is also a warning embedded in the price war. Commentators cited in the October 26 coverage split between "race to the bottom, someone goes bankrupt" and "this is exactly why both labs are suddenly so interested in regulation." A market where frontier intelligence gets 80% cheaper in four months is a market where today's cost advantage is not a moat. Founders who build their pitch on cheap model access alone are building on ground that moves every quarter.

The deeper opportunity is the one Apollo's Torsten Slok flagged in the same week: so far, the money has gone to companies selling AI, not buying it. Falling inference prices are what could change that — if founders use them to rebuild workflows rather than just to run the same workflows more cheaply.

Angeled.com connects founders with investors.

Source: minimaxir.com
Max Woolf (minimaxir), "OpenAI and Anthropic cut their LLM prices by up to 80% within a few months. Why?" (October 2026)
Source: ai0.news
AI0 News, "AI News — October 26, 2026: 80% Inference Price Cuts Traced to DeepSeek Pressure"
Source: news.lavx.hu
LavX News, "OpenAI and Anthropic slash LLM prices up to 80% as open-weight competition intensifies"
Source: wsj.com
Wall Street Journal, "The AI Price War Is Heating Up — and OpenAI Is Gaining Ground on Anthropic"

Read Next