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

OpenAI and Anthropic sharply reduced prices for their leading language models between June and September 2026, while releasing newer models that generally matched or improved on earlier performance. OpenAI cut its small Luna model by 80%, later halved prices for its GPT-6 models, and reduced GPT-6.1 Sol’s cache-reading charge. Anthropic made Sonnet 5’s launch discount permanent and lowered Opus 5.5’s overall costs by 40-60%. For coding agents, cache reads dominate usage because each turn reuses the conversation history; consequently, a lower cache price can make a model substantially cheaper in practice. Vercel data also shows token costs falling, even as total spending rises because people use increasingly capable models more.

The central explanation is competitive pressure, not simply temporary discounts to attract customers. Models are easy to swap into agent tools, and open-weight alternatives such as DeepSeek and Kimi are improving quickly while being served by many competing inference providers. With low switching costs and multiple suppliers, providers have reason to keep lowering prices rather than raise them to some presumed “true” cost, provided serving remains profitable. OpenAI and Anthropic must sustain a quality advantage and are building other defenses, including enterprise offerings and personal assistants. The timing may also relate to both companies’ rumored IPO plans: reported finances show large losses, so investors may scrutinize how they will become profitable. The price cuts therefore reflect both a competitive market and the business pressures surrounding their growth.