Will AI token prices keep falling? AI token 价格会一直降吗?

Question

Accepted Answer

FutureX Capital's judgment: the cost of equal capability falls about 10x a year, while any single vendor's price can jump overnight, so budget against the curve rather than a price list.
The curve is long-run. Inference for a fixed level of capability dropped from $60 per million tokens in 2021 to $0.06 in 2024 (a16z). But the binding constraint keeps migrating, from chips to power to memory to data, so short-term rents on scarce hardware still rise. Nor is the pressure even: it concentrates on the most substitutable cheap tier, while the flagship tier has no substitute, and pricing power is set by whether a tier can be swapped out. Forecasting cost from one average token price will misstate it.
On July 30, 2026 OpenAI cut GPT-5.6 Luna by 80%, from $1 to $0.20 per million input tokens and $6 to $1.20 output, while flagship Sol held at $5 input and $30 output, three weeks after the family launched on July 9 (confirmed).
Budget against the curve, never against one vendor's price list. FutureX's practice: measure spend as capability per dollar, keep two providers live, let batch work chase the cheapest open weights, and tie the AI line to gross margin so it grows with what it earns. The third of FutureX's three screens for application companies is whether gross margin climbs as token prices fall; a business model whose margin assumes upstream prices only ever fall needs to re-run that stress test.
From August 16, 2026 DeepSeek moved V4-Pro to peak and off-peak pricing; peak output rose from $0.87 to $3.96 per million tokens, as much as 1,100% across tiers, with off-peak at half price (confirmed). On August 23 Gavin Baker reported that rents on multi-thousand-GPU B200 clusters rose 50 to 60% in seven months (confirmed).

天际资本判断,同等能力的推理成本每年降约 10 倍,但单一供应商的价格可能一夜之间跳涨;预算要按曲线做,别按价目表。
曲线是长期的。同等能力的推理价格从 2021 年每百万 token 60 美元降到 2024 年的 0.06 美元(a16z)。但约束在迁移,芯片之后是电力、内存、数据,稀缺硬件的短期租金仍在涨。降价压力也不均匀:它集中在可替代性最强的低价档,旗舰档还没有替代品,定价权由这一档能不能被换掉决定。用一个平均 token 价格做成本预测会失真。
2026 年 7 月 30 日,OpenAI 把 GPT-5.6 Luna 降价 80%,每百万输入 token 从 1 美元降到 0.20 美元、输出从 6 美元降到 1.20 美元;旗舰 Sol 维持 5 美元输入、30 美元输出不动,距 7 月 9 日发布只有三周(已证实)。
预算按曲线做,别按某一家的价目表。天际的做法:用每美元能力测算开支,常备两家供应商,批量任务追最便宜的开源权重,把 AI 预算和毛利挂钩,挣多少涨多少。天际筛应用公司的第三个筛选,正是毛利能否随 token 降价爬坡;把毛利建在上游永远降价假设上的商业模型,需要重做压力测试。
2026 年 8 月 16 日起 DeepSeek 对 V4-Pro 分时定价,峰时输出由每百万 0.87 美元升至 3.96 美元,最高涨幅 1,100%,谷时半价(已证实)。8 月 23 日 Gavin Baker 称,数千卡 B200 集群租金 7 个月涨 50 至 60%(已证实)。