OpenAI and Anthropic have repeatedly cut large language model (LLM) usage fees from the summer through autumn of 2026, driven by low user switching costs and price competition with open-weight models.

On Oct. 29, online media outlet Gigazine reported that data scientist Max Woolf recently said the wave of price cuts shows a market structure in which it is difficult to lock users into a specific model.

OpenAI cut the price of its low-cost model GPT-5.6 Luna by 80 percent on July 30. The input fee per 1 million tokens fell to $0.20 from $1, and the output fee fell to $1.20 from $6. OpenAI later released material comparing the cost efficiency of GPT-5.6 Luna with open-weight models including DeepSeek V4 Pro and GLM-5.2.

Fees for its flagship model also fell as the generation changed. GPT-5.6 Sol, released in June, was priced at $5 per 1 million tokens for input and $30 for output. By contrast, GPT-6.1 Sol, unveiled on Sept. 29, was set at $2 for input and $10 for output. The cache read fee for reusing processed input is $0.10 per 1 million tokens.

Anthropic also adjusted pricing on Sept. 22 as it announced Claude Opus 5.5. It is priced at $4 per 1 million tokens for input, $20 for output and $0.20 for cache reads. Anthropic said it cut input and output fees by 20 percent from Opus 5 and reduced the cache read fee by 60 percent. It said cache reads account for a significant share of costs in AI agent tasks and coding, and that total costs for typical tasks are 40 percent lower than Opus 5.

A key driver of the price cuts is lower model switching costs. Woolf said models can be easily switched when using programming assistant tools such as OpenCode and Zed Agent. As the time and cost of moving to another model fall, it has become harder for a specific service to secure users over the long term.

Actual usage data also showed the trend. Vercel AI Gateway said 48.0 percent of teams switched their flagship model during September, and 28.6 percent switched model developers as well. On Hacker News, users shared accounts saying they used Kimi and GLM after OpenAI and Anthropic refused requests due to cybersecurity issues, and that costs were much lower and they were able to complete tasks to the end.

Competition with open-weight models also increased pressure to cut prices. Open-weight models disclose trained weights, allowing providers other than the developer to offer the same model. When multiple providers offer the same model, users can compare prices and choose cheaper services. As of Oct. 24, DeepSeek V4.1 Flash 1017 offered on OpenRouter had 29 providers. DeepSeek's own prices were $0.15 per 1 million tokens for input, $0.60 for output and $0.003 for cache reads, with the cache read fee about 2 percent of the regular input fee.

Token usage and spending patterns have also changed. Vercel AI Gateway data showed the token usage index in September rose to 790.8 versus April, while the spending index was limited to 342.7. The average token unit price fell to 43.3. Total spending still rose as usage increased faster despite the drop in unit prices. Open-weight models handled 71.5 percent of total processed tokens that month, but accounted for only 23 percent of costs. That shows cheaper alternative models are being used for large-scale tasks.

Lower supply costs from technical improvements were also cited as a factor behind the price cuts. Anthropic said Claude Opus 5.5 can be provided with fewer computing resources than Opus 5. OpenAI also said it has improved efficiency in its model and inference systems and external tool integration, enabling it to process more tasks with the same computing resources.

Woolf said in this market environment it has become difficult for OpenAI and Anthropic to secure users and then sharply raise prices to recoup profits. If multiple providers offer similar-performing models at low prices, users can move to other services instead of accepting price hikes.