OpenAI 于7月27日至8月14日期间，对其新的 Terra 和 Luna 模型推出了大幅折扣。这些折扣对 token 用量、总消费额以及竞争格局产生了怎样的影响？
要点概览
- Token 用量激增：折扣期间，Terra 模型每日 token 使用量增长了 5.6 倍，而 Luna 模型每日 token 使用量则跃升了 13.8 倍。同期，保持原价的 Sol 模型仅出现 1.1 倍的温和增长。
- 份额重新洗牌：OpenAI 折扣所获得的份额增长，大部分来自其他实验室，而非对 OpenAI 自家模型体系内部的蚕食。
- 用户留存显著：在折扣期间尝试过 OpenAI 折扣模型的用户中，近三分之一在折扣结束后仍继续使用该模型。
折扣对 token 使用量的影响
GPT 5.6 折扣一经生效，token 用量便呈爆发式增长。
如果将活动期间的 token 使用量与活动前的日均水平进行比较，效果一目了然：Terra 的 token 用量增长了 5.6 倍，Luna 增长了 13.8 倍。未参与折扣的 Sol 模型仅出现 1.11 倍的轻微增长，而 OpenAI 其他模型的 token 用量实际上略有下降。在 OpenAI 体系之外，其他模型则呈现温和增长。
竞争性替代效应
从活动前到活动期间，Terra/Luna 在 OpenRouter 全部 token 中的占比从 0.7% 上升至 7.8%，份额增长了 7.1 个百分点。所有竞争对手被合并到浅灰色区域中，每个色块上方的标签为 OpenAI 家族（Terra + Luna + Sol + 其他 OpenAI 模型）的合计份额。
在同一对比区间内，竞争对手让出了 5.3 个百分点的份额，OpenAI 其他模型让出了 1.9 个百分点，因此约四分之三的增长来自 OpenAI 体系之外。就整个 OpenAI 家族而言，token 份额从 7.1% 增长至 12.4%，且在折扣期间的某些特定日期一度突破 15%。
主要模型厂商的 token 用量大多有所上升，不过 Anthropic 在此期间并未增长。OpenAI 的 token 用量平均几乎翻了一番，并且在折扣活动结束后仍维持在这一较高水平。
用户留存情况
显然，折扣推动了活动期间 token 使用量的增长，但折扣结束后，这些用户是否留了下来？
在该计划期间使用过 Terra/Luna 的 10 万多名客户中，约有 32% 在后续几天仍保留了一定使用量，18% 的使用量达到或超过了其计划期间的水平。需要说明的是，这是客户数量统计，而非按 token 加权。显然，计划结束后的时间段比完整折扣期要短得多（目前为 6 天，而计划期为 19 天），因此随着更多数据的到来，情况可能会发生变化。
如果改看每日 Terra/Luna token 量，计划结束后的日均 token 量是折扣期间的 1.38 倍，这意味着留存下来的账户规模远大于计划用户的中间水平。
Sol 在 Terra/Luna 计划期间日均处理 791 亿 token，而计划前为日均 712 亿，基本是一个平稳的对照组。
不过，上图中从 8 月 17 日起的阴影区域是 Sol 自己的 50% 折扣。Sol 立即出现跳升，重现了 Terra/Luna 的模式。
方法论与说明
数据详情
- 计划前：7 月 8 日至 7 月 26 日 | 计划期：7 月 27 日至 8 月 14 日 | 计划后：8 月 15 日至 8 月 20 日
- 7 月 30 日：OpenAI 在通过 OpenRouter 提供 50% 折扣的基础上，又下调了自身目录价格（Luna 降 80%，Terra 降 20%）。因此，自 7 月 30 日起的实际折扣为 Luna 90%、Terra 60%。
- 说明：Terra、Luna 和 Sol 均于 7 月 9 日上线。Sol 在 8 月 17 日获得自己的 50% 折扣之前是一个有效的对照组。
方法论
- 排除项：被封禁、已删除、管理员、申请删除数据、内部及已流失的账户均被剔除。
- 8 月 20 日是统计窗口的最后一天，数据可能不完整。
- Sol 仅在 8 月 16 日之前是有效的对照组。计划后时间段为 6 天。
OpenAI introduced large discounts on their new Terra and Luna models from July 27th through August 14th. What impact did these discounts have on token volumes, total spend, and the competition?
Highlights
- Tokens spiked: Daily Terra token usage rose 5.6x during the discount window, while daily Luna token usage jumped 13.8x. The Sol model, which remained a list price, saw only a gentle bump of 1.1x over the same period.
- Share was reconfigured: Most of the share gained by the OpenAI discounts came from other labs as opposed to cannibalization within the OpenAI family of models.
- Users stuck around: Nearly a third of users who tried a discounted OpenAI model during the discount window kept using it after the discounts expired.
Discount impact on token usage
The moment GPT 5.6 discounts kicked in, the token volume exploded.
If you compare the in-program token usage to the pre-period daily averages, the effect is crystal clear: Terra tokens rose 5.6x and Luna 13.8x. The Sol model, which remained un-discounted, saw a minor 1.11x bump while other OpenAI models actually fell slightly in tokens used. Outside the OpenAI family, the other models rose gently.
Competitive displacement
Terra/Luna went from 0.7% to 7.8% of all OpenRouter tokens between the pre-period and the program, a gain of 7.1 share points. All competitors are pooled into the light-grey segment, and the label above each color stack is the combined OpenAI-family share (Terra + Luna + Sol + other OpenAI).
Competitors gave up 5.3 points and other OpenAI models gave up 1.9 points over the same comparison, so roughly three quarters of the gain came from outside OpenAI. Across the whole OpenAI family, token share grew from 7.1% to 12.4% and occasionally crested over 15% in specific days during the discount period.
Tokens mostly rose across key model authors, though Anthropic did not in this timeframe. OpenAI token volume nearly doubled on average and has maintained that higher level after the discount program ended.
Users retention
Clearly the discounts drove increased token usage during the program period, but did those users stick around after the discounts had ended?
Of the 100K+ customers with Terra/Luna usage during the program, about 32% retained some usage in the subsequent days and 18% ran at or above their program pace. To be clear, this is a count of customers, not weighted by tokens. Obviously the post-program period is quite a bit shorter than the full discount timeframe (6 days so far vs a 19-day program), so the story may change as more data comes in.
If you look instead at the daily Terra/Luna token volume, the post-program period saw 1.38x the tokens of the discount period on average, which implies the retained accounts are far larger than the median program user.
Sol averaged 79.1B tokens/day during the Terra/Luna program against 71.2B/day pre-period, an effectively flat control group.
However, the shaded region in the chart above from Aug 17 onward is Sol’s own 50% discount. Sol jumps immediately, reproducing the Terra/Luna pattern.
Methodology & Notes
Data Details
- Pre: Jul 8 to Jul 26 | Program: Jul 27 to Aug 14 | Post: Aug 15 to Aug 20
- Jul 30: OpenAI cut its own list prices (Luna 80%, Terra 20%) on top of the 50% discount through OpenRouter. As such, effective discounts from Jul 30 were Luna 90% and Terra 60%.
- Notes: Terra, Luna and Sol all launched Jul 9. Sol is a useful control up until it got its own 50% discount on Aug 17.
Methodology
- Exclusions: banned, deleted, admin, data-deletion-requested, internal and churned accounts are dropped.
- Aug 20 is the last day in the window and may be partial.
- Sol is a valid control only through Aug 16. The post-period is 6 days.