OpenAI and Anthropic cut their LLM prices by up to 80% within a few months. Why?

What it claims

Max Woolf argues that OpenAI and Anthropic list prices fell by 40-80% between June and September 2026 without any visible loss of output quality. The cuts apply to API use and to subscription harnesses such as Codex and Claude Code. Woolf puts the cause in three places: models that are hard to tell apart and easy to swap, price pressure from open-weight models served by many third-party hosts, and possibly IPO positioning. Woolf also believes Anthropic's statement that Opus 5.5 needs less compute to serve than Opus 5. Woolf rejects the claim that token prices are kept artificially low and will later jump. The post's view is that third-party hosts drive prices toward marginal cost, because they stop serving any model that loses money.

Evidence

All prices are input / output / cache read, per million tokens.

• OpenAI: GPT-5.6 Luna launched on June 26 at $1 / $6 / $0.10. It was cut on July 30 to $0.20 / $1.20 / $0.02. GPT-5.6 Sol launched at $5 / $30 / $0.50 and was temporarily cut on August 21 to $4 / $20 / $0.40 for three months. GPT-6 Astra launched on September 3 at $10 / $50 / $1. GPT-6 Sol launched on September 22 at $2 / $10 / $0.20. Woolf calls it "a dud" on performance. GPT-6.1 Sol followed on September 29 at $2 / $10 / $0.10, and Woolf reports from personal use that it performs close to Astra. GPT-6 Luna is $0.10 / $0.50 / $0.01.
• Anthropic: Sonnet 5 had launch pricing of $2 / $10 / $0.20, made permanent on August 10. Opus 5 launched at $5 / $25 / $0.50. Fable 5.1 launched on September 1 at $10 / $50 / $0.25, against $1 cache read for Fable 5. Opus 5.5 launched on September 22.

The cache figure drives the effective price. In coding agents, cache reads are 96-99% of tokens and hit 90-100% of the time. So a 75% cut to cache read price gives roughly a 70-75% cut in total cost. Opus 5.5's 20% input/output cut plus its larger cache cut gives a 40-60% total reduction against Opus 5.

On open weights, DeepSeek V4 Flash 0420 (April 2026) cost $0.14 / $0.28 / $0.0028. That cache read is 2% of the input price, against 5-10% for the OpenAI and Anthropic models. DeepSeek Flash V4.1 1017 scores 45 on the Artificial Analysis index, against 24 for V4. DeepSeek prices it at $0.15 / $0.60 / $0.003. It has 29 OpenRouter providers, against 13 for V4 Flash in May, and many sit close on effective input cost.

Kimi K3.1 Pro scores 55 on the same index, against 58 for Opus 5.5. It costs $2 / $10 / $0.20, down from Kimi K3's $3 / $15 / $0.30. Its weights are due on November 4. Kimi K3 already has more than 20 hosts, so Woolf expects similar third-party pricing for K3.1 Pro.

How to use it

For agent workloads, compare models on cache read price and provider cache hit rate, not list input price. OpenRouter requests can set a "lowest price" routing preference.

Subscriptions are moving the other way. When GPT-6.1 Sol launched, OpenAI cut the Codex $200/month tier's usage allowance by 50%. Woolf expects subscription value to be cut before base model prices if capital gets tight.

Limits

The IPO reasoning is speculation: OpenAI is rumored for 2027 and Anthropic within a couple of months. The 80% inference margin figure comes from unnamed reports "with caveats." The quality judgments rest on Woolf's own use and on benchmarks, not controlled tests. The post gives no per-provider OpenRouter prices or cache hit rates.