OpenAI and Anthropic Keep Cutting LLM Prices, Some by Up to 80%

OpenAI and Anthropic have cut the prices of their flagship language models repeatedly since June, and one cut reached 80 percent. A detailed timeline published by an independent analyst tracks the moves, which began with a major OpenAI launch in late June and have continued into the fall.

The sequence started on June 26, when OpenAI announced GPT-5.6 Sol and Luna. Sol replaced GPT-5.5 at the same price, $5 per million input tokens and $30 per million output tokens. Luna was a smaller model aimed at cheaper tasks. Four days later, Anthropic launched Claude Sonnet 5 with introductory pricing of $2 and $10, a rate the company had initially described as temporary. On August 10, Anthropic made that lower price permanent.

The biggest single cut came on July 30. OpenAI said it was reducing the price of GPT-5.6 Luna by 80 percent, to $0.20 per million input tokens and $1.20 per million output tokens. According to the analysis, the cut was aimed at small open-weight models that had been drawing developers with low prices. Without it, the writer argues, Luna would have lost to DeepSeek V4 Pro on cost per task.

The pace picked up in the fall. On August 21, OpenAI cut GPT-5.6 Sol by about 20 percent for three months. On September 1, Anthropic released Fable 5.1 at $10 and $50 and cut its cache read cost by 75 percent. Cache reads account for most of the spending when developers run AI agents, the analysis says, because the cache holds the entire conversation on every turn. That means a cut in cache pricing can translate into a large drop in what a user actually pays.

On September 22, OpenAI launched GPT-6 Sol at $2 and $10, half the price of the temporarily discounted GPT-5.6 Sol, and released a cheaper GPT-6 Luna on the same day. Anthropic announced Claude Opus 5.5 that day as well. The company said in its announcement that Opus 5.5 requires less compute to serve than Opus 5, and that its pricing reflects that. The analysis estimates the total cost of using Opus 5.5 is 40 to 60 percent lower than Opus 5. A week later, OpenAI released GPT-6.1 Sol, cutting its cache price in half to $0.10 per million tokens.

The analysis says the timing suggests the companies are reacting to each other, though it stops short of confirming that. It also points to competition from outside the two labs. Models have become easier to swap, the writer says, because developers can test a new model in a coding tool or through a single API endpoint. Open-weight models from providers including DeepSeek and Moonshot AI are closing the gap on performance. DeepSeek V4.1 Flash, for example, scored 45 on the Artificial Analysis index, up from 24 for the earlier V4 Flash, while keeping a similar price from DeepSeek itself.

Many companies now host those open-weight models. On OpenRouter, a platform that routes requests to different providers, the number of providers offering DeepSeek V4.1 Flash 1017 has grown to 29, up from 13 for the earlier version. The analysis found that several of them have effective input costs close to DeepSeek's own. Moonshot AI's Kimi K3.1 Pro, announced shortly before the analysis was published, is priced at half the input and output cost of Opus 5.5 and has a performance score the writer describes as comparable, 55 against 58.

Both companies are also preparing to go public. The analysis says OpenAI is rumored to be planning an IPO in 2027, and Anthropic within a couple of months. Leaked financial reports from both companies show large losses tied to research and scaling, though the writer notes those reports may not reflect the recent surge in agent usage. Cutting prices on a main revenue source could look odd in an IPO filing, the analysis suggests, and the profit margins on inference will be a key figure investors look at.

Cost cuts have not always reached subscribers in the same way. OpenAI cut the usage allowance by 50 percent for its $200-a-month Codex tier at the same time it launched GPT-6.1 Sol, saying it reflected more efficient models. The writer, who said he had been using Codex, said he cut his own usage limit after the price fell, leaving his monthly cost roughly unchanged.

Vercel's AI gateway data, cited in the analysis, shows cost per token has fallen to 43.3 percent of its April 2026 level. Total spending on AI models has still risen, the writer notes, because the models are improving faster than their prices are falling.

Este artículo fue escrito con la asistencia de IA.
News Factory APP - noticias agénticas para impulsar tu SEO y AEO.