OpenAI 与 Anthropic 数月内大模型降价高达 80%，背后原因何在？

让我们快速梳理一下过去几个月来自 OpenAI 和 Anthropic 的主要 LLM 模型发布时间线。下文提到的所有定价均为每百万 Token 的 输入 / 输出 / 缓存读取（input MTok / output MTok / cache read MTok）。

2026 年 6 月 26 日#

OpenAI 发布了 GPT-5.6 Sol 和 Luna。Sol 是 GPT 5.5 的直接继承者，价格同样维持在 $5 / $30 / $0.50；而 Luna 则是专为低成本场景设计的小模型，定价为 $1 / $6 / $0.10。他们还发布了一款介于两者之间的 Terra 模型，但鲜有人关注。

2026 年 6 月 30 日#

Anthropic 发布了 Claude Sonnet 5，Anthropic 声称其能力更接近其 Opus 系列模型，但外界反响褒贬不一。值得注意的是，其首发定价为 $2 / $10 / $0.20，原本计划为限时价格、后续会上调。

2026 年 7 月 24 日#

Anthropic 发布了 Claude Opus 5。鉴于 Opus 4.5 对 Agent 编程兴起带来的深远影响，这是一个备受期待的大版本更新，定价为 $5 / $25 / $0.50（与 GPT-5.6 Sol 相仿）。尽管其公布的部分基准测试结果令人意外地超越了体积大得多的 Fable 5，但在实际应用中的反馈却更为复杂。

2026 年 7 月 30 日#

OpenAI 宣布将 GPT-5.6 Luna 的价格直降 80%（？？？）至 $0.20 / $1.20 / $0.02。

大幅降价后 GPT-5.6 Luna 的模型质量与单任务成本对比。越靠左上方越好，如果 OpenAI 不降价 80%，它就会输给 DeepSeek V4 Pro。来源：OpenAI

2026 年 8 月 10 日#

Anthropic 宣布 Sonnet 5 的首发限时优惠转为永久定价。

2026 年 8 月 21 日#

OpenAI 宣布在接下来的三个月里，将 GPT-5.6 Sol 的价格下调约 20%（$4 / $20 / $0.40）。

2026 年 9 月 1 日#

Anthropic 发布了 Fable 5.1，价格为 $10 / $50 / $0.25。除了在各项基准测试中全面优于 6 月份发布的 $10 / $50 / $1 的 Fable 5 之外，还将缓存读取费用削减了 75%。

在利用 Agent 编写代码时，缓存读取成本占据了总 Token 支出的绝大部分（96% - 99%），因为每一轮交互都需要在缓存中加载完整的历史对话。此外，缓存命中率通常在 90% 到 100% 之间（视 LLM 推理服务商的基础设施而定）。综合来看，缓存读取费用下降 75%，直接等同于模型调用总成本下降 70-75%——即使缓存费用在报价中通常只是个不起眼的脚注，但这显然是极其重大的变化！

2026 年 9 月 3 日#

OpenAI 发布了 GPT-6 Astra，其体积与 Fable 5.1 相当，并采取了类似的 $10 / $50 / $1 定价，但没有提供较低的缓存读取成本。Astra 的模型性能确实可以与 Fable 相抗衡。

2026 年 9 月 22 日#

OpenAI 发布了 GPT-6 Sol，开出了令人震惊的 $2 / $10 / $0.20 定价，相较于此前限时折扣的 GPT-5.6 Sol，其输入和输出价格直接腰斩！他们还同时推出了 GPT-6 Luna，同样以旧版一半的输入输出价格（$0.10 / $0.50 / $0.01）发售。

就在同一天，Anthropic 发布了 Claude Opus 5.5，并明确表示：“Opus 5.5 运行所需的算力低于 Opus 5，其定价也反映了这一点。”由于输入/输出成本降低 20%，加之缓存读取成本大幅下降，其 $4 / $20 / $0.20 的定价在总体使用成本上远低于 Opus 5；对于一款性能甚至优于 Fable 5.1 的模型而言，这相当于总体成本降低了 40-60%（关于 Opus 5.5 我有很多要谈的，但那要留到另一篇博客了）。

2026 年 9 月 29 日#

事实证明 GPT-6 Sol 在性能上是个半成品，因此一周后 OpenAI 发布了 GPT-6.1 Sol。该模型的性能出人意料地达到了 Astra 的水平（经过个人实际使用确认无误），但 OpenAI 顺便将缓存价格也砍半至 $2 / $10 / $0.10，与 Anthropic 缓存读取仅占输入费用 5% 的比例看齐，使得这一模型系列累计降价达到 75%。

总体而言，从 6 月份发布 GPT-5.6 系列模型至今，调用这些主流大模型的费用发生了断崖式下跌，这是一年前我完全无法预料的。这一波降价也同步覆盖了 Codex 和 Claude Code 等订阅式 Agent 工具。更关键的是，这些模型在降价的同时，似乎并没有在输出质量上偷工减料。

来自 Vercel AI 模型网关的数据支持了单 Token 成本降至 2026 年 4 月成本 43.3% 的趋势。然而，总支出却一直在增加，因为模型变好的速度超过了它们降价的速度！来源：Vercel

从上面的时间线可以看出，Anthropic 和 OpenAI 之间存在着清晰的针锋相对，有理由推测双方的这些降价动作是在互相见招拆招。然而，若没有外部因素的作用，没人会主动将价格下调 80%。那么，真正的原因到底是什么？

大模型的行为经济学与竞品定价压力#

如果你一直紧跟新模型发布并在 Agent 场景中进行测试，就会发现除了主观的“感觉”（vibes）之外，要在它们之间做出区分已经变得非常困难。自去年 11 月引爆 Agent 大模型爆发式增长以来，模型正愈发成为一种大宗商品化（commodity）的存在：在 OpenCode 和 Zed Agent 等工具中只需轻轻一按切换开关，或者在 OpenRouter 中更换一行不同的 API 端点字符串，就可以毫无门槛地在代码库上试用新模型，因此护城河微乎其微。此外，还存在大量竞争对手——不仅有 Google 和 Meta 这样的大厂——他们都能提供构成充分替代品的模型，从而形成了强劲的定价压力。

在开放权重模型方面，我们再来看看 DeepSeek Flash V4.1，该模型最近也推出了更新的 1017 变体。

目前来看，OpenRouter 是使用开放权重模型的首选之地。获取于 2026 年 10 月 24 日。

最初的 DeepSeek V4 Flash 0420 于 2026 年 4 月发布，价格为 $0.14 / $0.28 / $0.0028，并迅速引爆全网，特别是在当时 OpenAI 还没有针对简单任务的小型模型的情况下。事实证明，那个 $0.0028 成了胜负手：仅占 2% 的缓存读取比例远低于 OpenAI 和 Anthropic 上述模型的 5%-10%，这使得实际综合使用成本比广告宣称的还要低。我推测 GPT-5.6 Luna 在 7 月份降价 80%，正是为了针对性地同这些较小的开源权重模型竞争。如今，新的 DeepSeek Flash V4.1 1017 模型在基准测试中的表现显著优于 DeepSeek Flash V4（在 Artificial Analysis 指数上得分为 45，而之前为 24），而来自 DeepSeek 官方的成本依然保持在接近的 $0.15 / $0.60 / $0.003。请注意关键词：“来自 DeepSeek 官方”。

在 OpenRouter 上，任何推理提供商都可以上架自己托管的模型。当用户请求 OpenRouter API 时，请求会被路由到其中一个模型实例——默认情况下为了负载均衡会路由到任意可用模型，但开发者也可以在 API 请求中指定偏好，例如“最低价格”。早在 5 月份我就指出，在 OpenRouter 上托管 DeepSeek V4 Flash 的独立服务商就有 13 家。而对于更新后的 1017 变体，如今已有 29 家不同服务商。

按缓存读取成本排序的 DeepSeek V4.1 Flash 1017 服务商列表。获取于 2026 年 10 月 24 日。

然而，观察实际有效输入成本（综合考量了缓存读取成本与缓存命中率之间的相互影响），许多不同服务商的成本其实非常接近。

各端点的有效成本，其中 Cloudflare 和 DeepInfra 与 DeepSeek 官方成本的差距在 10% 以内。尽管这些模型是最便宜的，但请注意，这些模型仅占总 Token 份额的一小部分。获取于 2026 年 10 月 24 日。

OpenAI 和 Anthropic 必须依靠更高质量的模型来与这些开放权重模型建立区隔——他们目前在这方面做得很好，尤其是 Opus 5.5——但他们必须持续保持领先，这也是数十亿美元不断投入研发的原因。开源模型正在迎头赶上：就在几天前，月之暗面（Moonshot AI）发布了 Kimi K3.1 Pro，其性能可与 Opus 5.5 相媲美（Kimi 得分为 55，Opus 5.5 为 58），输入/输出价格却只有后者的一半，且具有相同的缓存读取成本（$2 / $10 / $0.20）。它的尺寸和架构与 Kimi K3 完全相同，后者目前由 20 多家不同的推理服务商托管：因此，一旦权重在 11 月 4 日发布，你可以预见其中大多数服务商都会以大致相当的价格支持新的 Kimi K3.1 Pro 模型。值得注意的是，Moonshot AI 本身也加入了降价狂欢，Kimi K3.1 Pro 的定价甚至低于初代 Kimi K3（$3 / $15 / $0.30）。

AI 批评者中经常出现的一种论调是：Token 的价格一直被刻意压低以吸引人们进入 LLM 生态系统，但未来价格终将上涨至“Token 的真实成本”，一旦“免费便车”到头，AI 狂热者们就会傻眼。该理论认为 OpenAI 和 Anthropic 正在采用亏本倾销（loss-leading）策略，就像 Uber 和 Lyft 在 2010 年代对网约车行业所做的那样。然而与当年不同的是，OpenAI 和 Anthropic 并不是拥有全行业唯一定价权的寡头，推理行业的准入门槛也没有那么高。只要这众多的推理提供商能够继续以与 OpenAI 和 Anthropic 相似的价格和质量提供模型服务，并且开发者的切换成本微乎其微，那么价格就可能继续进一步走低。不难理解两家公司都在试图建立护城河来抵御这一趋势，例如重点发力企业级承诺协议以及像 OpenAI Dots 和 Claude Duet 这样的个人 Agent 助手。

认为 Token 价格会突然暴涨的想法更像是一种幻想，既没有行为经济学的支持，也不符合过去几个月 LLM 在模型质量提升的同时价格大幅走低的现实趋势，而且降价的远不止 OpenAI 和 Anthropic。只要规模化扩张者的每个 Token 成本低于每个 Token 的营收，这些推理商就会为了互相竞争而不断降低 Token 成本，直到迫使 OpenAI/Anthropic 也不得不降价以保住竞争力。OpenRouter 上的 29 家不同服务商几乎没有动力去托管一个赔钱的模型（他们完全可以停止服务并把算力分配给能盈利的模型），而“Token 的真实成本”最终可能会稳定在价格等于边际成本的平衡点。

说到这里，OpenAI 和 Anthropic 确实都临近 IPO，这引发了关于降价动机的更多疑问。

为 IPO 造势#

这些降价的时机之所以引人注目，是因为传闻 OpenAI 和 Anthropic 都即将进行 IPO——[OpenAI 预计在 2027 年，而 Anthropic 预计在几个月内](https://www.investing.com/news/stock-market-news/e

Let’s look at a quick timeline of the major LLM model announcements from both OpenAI and Anthropic over the past few months. All pricing numbers provided are input MTok / output MTok / cache read MTok.

June 26, 2026#

OpenAI announced GPT-5.6 Sol and Luna. Sol is the direct successor to GPT 5.5 at the same price of $5 / $30 / $0.50, while Luna is a small model specifically designed for cheaper use cases at $1 / $6 / $0.10. They also announced a Terra model that’s in between those two models, but few cared about it.

June 30, 2026#

Anthropic announced Claude Sonnet 5, which Anthropic claimed to be closer to their Opus models, but feedback was mixed. Notably, it had launch pricing of $2 / $10 / $0.20 which was initially supposed to be temporary but would go up.

July 24, 2026#

Anthropic announced Claude Opus 5, which was a long-awaited major-version update given the impact Opus 4.5 had on the rise of agentic coding, at a price of $5 / $25 / $0.50 (similar to GPT-5.6 Sol). Although some of the provided benchmark results surprisingly exceeded those of the much larger Fable 5, real-world feedback was more mixed.

July 30, 2026#

OpenAI announces that they cut the price of GPT-5.6 Luna by 80% (???) to $0.20 / $1.20 / $0.02.

Model quality vs. Cost-per-task for GPT-5.6 Luna after the massive price reduction. Top-and-left is better, and if OpenAI didn’t do the 80% price cut, it would have lost to DeepSeek V4 Pro. via OpenAI

August 10, 2026#

Anthropic announces the temporary Sonnet 5 price cut for the launch is now permanent.

August 21, 2026#

OpenAI announces that they are cutting the price for GPT-5.6 Sol by ~20% ($4 / $20 / $0.40) for the next three months.

September 1, 2026#

Anthropic announced Fable 5.1 at a price of $10 / $50 / $0.25, which, in addition to having strictly better benchmark performance in all areas than Fable 5 from June at $10 / $50 / $1, cut the cache read cost by 75%.

When using agents for coding, cache read costs are the vast majority (96% - 99%) of the total token expenditure as the cache is what contains the entire previous conversation each turn. Additionally, the cache tends to be hit around 90 - 100% of the time (varies by the infrastructure of the LLM inference provider). Taken together, a 75% reduction in the cache read cost correlates to a 70-75% reduction in the cost of using the model—obviously notable, even if cache costs are typically a footnote!

September 3, 2026#

OpenAI announced GPT-6 Astra, which is similar in size to Fable 5.1 and has similar pricing at $10 / $50 / $1 but without the lower cache read cost. Astra’s model performance is indeed comparable to Fable.

September 22, 2026#

OpenAI announced GPT-6 Sol at at a shocking price of $2 / $10 / $0.20 that halves the input and output price relative to the temporarily-discounted GPT-5.6 Sol! They simultaneously released GPT-6 Luna also at half the input and output price ($0.10 / $0.50 / $0.01) of the previous version.

On the same exact day , Anthropic announced Claude Opus 5.5, where it’s explicitly stated “Opus 5.5 requires less compute to serve than Opus 5, and its pricing reflects that”. Its price of $4 / $20 / $0.20 is much lower in total than Opus 5’s due to a 20% input/output cost cut and a larger cache read cut, resulting in a total cost reduction of 40-60% for a model that’s even better than Fable 5.1 (I have quite a lot of things to say about Opus 5.5, but that’s for another post).

September 29, 2026#

It turned out that GPT-6 Sol was a dud performancewise, so a week later OpenAI announced GPT-6.1 Sol, which surprisingly has comparable performance to Astra (can confirm after personal use), but OpenAI casually cut the cache price in half ($2 / $10 / $0.10) matching Anthropic’s 5%-of-input cache read cost, making this model family cumulatively 75% cheaper.

Overall, starting from the release of the GPT-5.6 model families in June to now, the cost of using these popular LLMs has cratered, which is not something I would have expected a year ago. These price cuts also apply to model usage from subscription agentic harnesses such as Codex and Claude Code. More importantly, these models do not appear to have cut any corners in terms of output quality to achieve their price cuts.

Data from Vercel’s AI model gateway supporting the decreasing cost-per-token trend to 43.3% of the April 2026 cost. However, aggregate spend has been increasing because the models are getting better faster than they’re getting cheaper! via Vercel

From the timeline above, there is a clear tit-for-tat between Anthropic and OpenAI, and it is reasonable to assume that these particular price cuts from OpenAI and Anthropic are in reaction to each other. However, no one cuts prices by 80% unless there is some externality in play. So, what gives?

The Behavioral Economics of LLMs and Competitor Pricing Pressure#

If you have been testing agents with new LLMs as they are released, it can be very difficult to differentiate them outside of subjective vibes. Since last November that launched the hypergrowth of agentic LLMs, models have become more of a commodity where it’s trivial to try a new model on your codebase with just a toggle in a harness like OpenCode and Zed Agent, or a different API endpoint string through OpenRouter, and as a result there is little moat. Additionally, there are many competitors—not just large companies like Google and Meta—who can provide models which are sufficient substitutes and therefore allow pricing pressure.

For open weight models, let’s take another look at DeepSeek Flash V4.1, which also recently released an updated 1017 variant.

OpenRouter at this point is the place you go for open-weight models. Retrieved October 24th, 2026.

The original DeepSeek V4 Flash 0420 was released in April 2026 at a price of $0.14 / $0.28 / $0.0028 and went megaviral as a result, especially as OpenAI didn’t have a small-class model for simple tasks at the time. That $0.0028 turned out to be the key, as a cache read proportion of 2% is substantially lower than the 5%-10% of the OpenAI and Anthropic models above, which makes the effective cost even lower than what is advertised. I suspect the 80% GPT-5.6 Luna price cut in July was meant to specifically compete with these smaller open-weight models. Now, the new DeepSeek Flash V4.1 1017 model performs significantly better than DeepSeek Flash V4 benchmark-wise (45 on the Artificial Analysis index vs. 24) and the cost from DeepSeek remains similar at $0.15 / $0.60 / $0.003. Note the operative words “from DeepSeek”.

On OpenRouter, any inference provider can list a model that they host, and when a user hits the OpenRouter API, the request is routed to one of those models—it’s routed to any available model by default for load balancing but the developer can specify preferences such as “lowest price” in the API request. Back in May, I noted that there were 13 distinct providers hosting DeepSeek V4 Flash on OpenRouter. For the updated 1017 variant, there are now 29 distinct providers.

Providers for DeepSeek V4.1 Flash 1017 sorted by cache read cost. Retrieved October 24th, 2026.

However, looking at the effective input costs (which account for the interaction between cache read costs and cache hit rates), many distinct providers are close costwise.

Effective costs by endpoint, with Cloudflare and DeepInfra within 10% of DeepSeek’s cost. Despite the fact that these models are the cheapest, note that these models only have a minority of the total token share. Retrieved October 24th, 2026.

OpenAI and Anthropic will have to differentiate from these open weight models with higher-quality models—they are currently doing a good job with that particularly with Opus 5.5—but they need to keep doing that, which is why billions of dollars are being invested into R&D. Open-weight models are catching up fast: just a few days ago, Moonshot AI announced Kimi K3.1 Pro which has comparable performance to Opus 5.5 (Kimi at 55 vs. Opus 5.5 at 58) at half the input/output price and the same cache read cost ($2 / $10 / $0.20). It’s the same size and architecture as Kimi K3 which is currently being hosted by over 20 distinct inference providers: once the weights drop on November 4th, you can therefore expect most of them to support the new Kimi K3.1 Pro model at a roughly equivalent price. Notably, Moonshot AI itself is joining the price cut fun and priced Kimi K3.1 Pro at less than the original Kimi K3 ($3 / $15 / $0.30).

One recurring trend among vocal AI detractors is that the price of tokens has been kept artificially low to hook people into the LLM ecosystem, but then the prices will increase to “the true cost of a token” and the AI boosters will be screwed once the free ride ends. The theory is that OpenAI and Anthropic are loss-leading just like what Uber and Lyft did to the ridesharing industry in the 2010s, but unlike then, OpenAI and Anthropic are not a duopoly who have the sole power to set prices in the ecosystem and the barrier to entry for the inference industry is not as high. As long as these many inference providers can continue to serve models at a similar price and a similar quality to what OpenAI and Anthropic can provide, and switching cost is negligible, then the prices may continue to decrease more. It’s no surprise both companies are trying to build moats to defend against this trend, such as a focus on enterprise commitments and personal agentic assistants like OpenAI Dots and Claude Duet.

The idea that token prices will suddenly rise is more a fantasy that is not supported by both behavioral economics and the actual trends of LLM prices going down significantly over the last few months for better models, and not just from OpenAI and Anthropic. As long as scalers can price at a cost-per-token lower than the revenue-per-token, then those inference providers will reduce token costs to compete with each other to a point that OpenAI/Anthropic will have to lower their prices to compete too. The 29 distinct providers on OpenRouter have little incentive to serve an unprofitable model (they would just stop serving it and use the compute capacity for a profitable model instead) and “the true cost of a token” may end up being where price equals marginal cost.

On that note, both OpenAI and Anthropic are indeed close to IPOing, which raises further questions about the price cuts.

IPO Pumping#

The timing of these price cuts is notable because both OpenAI and Anthropic are rumored to be IPOing soon—OpenAI in 2027, and Anthropic within a couple months. So far, leaked financials from both OpenAI and Anthropic show unsurprisingly that they are losing billions due to R&D/scaling, albeit in both cases these reports do not account for the hyperexpansion of agentic LLM use over the past few months. Investors will scrutinize the massive negative profits and demand that they have a plan to turn things back around.

So, why did they cut their primary revenue source by up to 80%? That’s the opposite of what you would want to do to look good on an S-1 prospectus! The profit margins on inference themselves likely tell a different story, with some reports saying the profit-per-token margins are as healthy as 80% with caveats. How much of a profit margin do they have after these major cuts?

Speaking personally, the price cuts have indeed affected how I spend on LLM agents, but not in the way I expect OpenAI/Anthropic want their typical customer to behave. I used to have a $100/month subscription to Codex (the 5x Pro plan) when GPT-5.6 Sol was the top model, and I managed to just fully exhaust all the usage included in the subscription despite getting annoyed with random usage resets. Once GPT-6.1 Sol released and I noticed that the cost of using it was effectively cut by 75%, you know what I did?

I cut the max usage quota to 1/5th after getting 4x more efficient usage of the same model, so it ends up roughly neutral and the obvious choice.

In fairness, the profit margins for subscriptions such as Claude Code and Codex are uncertain given their immense value relative to the API usage, and if there are capital crunches I suspect subscription value will be cut before the base price of the models. For example, at the same time GPT-6.1 Sol was announced, OpenAI cut the usage allowance by 50% for the $200/month tier in Codex under the argument “to reflect our increasingly efficient models”—which is technically correct but it’s funny to put in writing.

I will be very curious to see the Anthropic IPO prospectus and the inference profit margins once it becomes public soon.

I also question whether it’s good for OpenAI/Anthropic if they intentionally obsolete their more expensive models. After Opus 5.5 released a month ago, there hasn’t been any discussion about either Fable 5.1 or GPT-6 Astra because Opus 5.5 is just that good.

Top-and-left. via Anthropic

I fully expect a Fable 5.2/GPT-6.1 Astra to release in a couple weeks and prove me wrong, and I will be very curious to see how it will be priced.

Another less-cynical factor that could explain the price cuts is that LLM infrastructure has improved over the past few months—when Anthropic said “Opus 5.5 requires less compute to serve than Opus 5, and its pricing reflects that,” I believe it. One outcome that would be genuinely ironic is that if these models are more compute efficient, it reduces the need for these massive data center buildouts, which could cause cascading economic failures. However, that likely won’t happen since inference is only one part of the compute mix and those data centers are still needed for other parts of the pipeline.

Overall, the price cut is weird, but as a user I’m certainly not complaining! Open-weight models and secondary inference providers will continue to serve as an economic check on the prices of LLMs and other business shenanigans the large labs may concoct. Even if Anthropic’s and OpenAI’s IPOs are catastrophic and cause both companies to fold, it will not suddenly cause LLMs and agents as an industry to die.