OpenAI and Anthropic model prices fell sharply since June: a timeline and why

Data scientist Max Woolf lists the price changes since June (all per million tokens). OpenAI cut GPT-5.6 Luna by 80% in July, to $0.20 input and $1.20 output; GPT-6.1 Sol now costs $2 input, $10 output and $0.10 for cached input, and Anthropic's Claude Opus 5.5 costs $4, $20 and $0.20, which he puts at 40 to 60% less in total than Opus 5. He says cached input makes up 96 to 99% of token spend for coding agents, so cache price cuts matter most there. He links the cuts to competition from cheap open-weight models served by many providers (29 host DeepSeek's latest Flash model on OpenRouter) and to cheaper serving. Vercel AI gateway data in the post shows cost per token down to 43.3% of April's level, while total spend keeps rising. On Hacker News (129 comments, 18 top-level), at least 6 top-level commenters put the cuts down to competition and a price war, 4 of them naming Chinese or open models. At least 5 pointed to cheaper serving instead, such as better caching, smaller models or more efficient hardware. The biggest thread (37 replies) began with one team saying they moved to Kimi and GLM models, which were cheaper and did not refuse their security work; replies turned to whether the US might ban Chinese models and which coding harness to use. No consensus: the competition camp and the efficiency camp both have support. Sorted take: prices are falling, but plans can shrink too. Woolf notes OpenAI halved the usage allowance on its $200 a month Codex tier when it launched GPT-6.1 Sol, so check both the token price and your plan limits.