OpenAI and Anthropic have cut AI model prices by as much as 80% since June

An analysis published on the blog minimaxir.com, widely shared on Hacker News, tracks a string of price cuts that OpenAI and Anthropic have made to their AI language models since June, with some reductions reaching 80%. The cuts touch three separate charges: the price per million input tokens, the price per million output tokens, and the price per million cached tokens, the portion of a saved conversation a model does not have to reprocess from scratch.

The clearest example is OpenAI’s GPT-5.6 Luna, a smaller model built for simple tasks. It launched priced at $1 per million input tokens, $6 per million output tokens, and $0.10 per million cached tokens. Months later OpenAI cut all three figures by 80%, to $0.20, $1.20, and $0.02. A chart OpenAI published comparing model quality to cost per task shows that without that cut, Luna would have lost on price to DeepSeek’s V4 Pro model, an open-weight competitor whose code is published for anyone to run.

Anthropic made comparable moves. It launched Claude Sonnet 5 at a temporary discount, then made that lower price permanent. Its newer Claude Opus 5.5 costs $4 per million input tokens, $20 per million output tokens, and $0.20 per million cached tokens, which Anthropic says reflects that Opus 5.5 “requires less compute to serve than Opus 5.” Combined with a larger cut to the cache-read price, that adds up to a 40-60% total cost reduction for a model Anthropic says performs better than Opus 5. OpenAI and Anthropic announced their respective cuts on the same day, continuing what the analysis calls a pattern of tit-for-tat pricing moves between the two companies dating back to June.

Why the prices keep falling

The analysis argues the cuts reflect competition rather than financial trouble. Coding tools now let developers switch between models with a single setting change, often through routing platforms such as OpenRouter, which makes it easy to compare models on price and hard for any one company to charge a premium.

Open-weight models add more pressure. DeepSeek’s V4 Flash, priced at $0.14 per million input tokens at its April launch, is now listed by 29 separate inference providers on OpenRouter, up from 13 in May, according to OpenRouter data cited in the analysis. Moonshot AI’s new Kimi K3.1 Pro reportedly scores close to Claude Opus 5.5 on the Artificial Analysis benchmark index while priced at half of Opus 5.5’s input and output cost.

Separately, data from Vercel’s AI model gateway cited in the analysis shows the cost per token across models has fallen to 43.3% of its April 2026 level, even as total spending on AI models keeps rising, because usage is growing faster than prices are falling.

The timing is notable because both OpenAI and Anthropic are reportedly preparing to go public, OpenAI in 2027 and Anthropic within a couple of months, and both have reported heavy losses tied to research and infrastructure spending. The source material does not detail current profit margins on inference itself. What happens to prices as more providers begin listing models like Kimi K3.1 Pro, due for release on November 4, is the next thing to watch.