Wink Pings

$200 Annual Fee, $100,000 Bill: How Does Anthropic Make Money?

Vlad Mihalcea posted a bill questioning Anthropic's business model. The answer lies in KV cache and how coding agents work: output tokens are only a small part; repeatedly reading context is the big part, and cache reads are almost free for the provider.

Vlad Mihalcea posted a screenshot of a bill on X: the Pro plan costs $200 a year, he used 16 billion tokens, and at API list prices that would be about $100,000. He asked how Anthropic's business model makes sense.

Gergely Orosz retweeted this and gave a very engineer-style explanation: inference engineering 101, reading KV cache is cheap, and speculative decoding makes sessions more efficient. Coding scenarios use both heavily, so inference is very profitable.

Gergely Orosz 推文截图

Vlad Mihalcea 账单截图

Let's fill in the background clearly. LLM inference has two steps. The prefill stage reads the input, runs the entire prompt through a forward pass, and the cost is related to the number of input tokens. The decode stage generates output tokens one by one; for each token generated, the model must read the entire context from high-bandwidth memory. Output tokens are more expensive, mainly because of memory bandwidth.

KV cache is an intermediate result. The model stores representations of context it has already processed, so the next time it encounters the same context, it doesn't have to recompute it. Reading the cache consumes almost no compute and doesn't use much memory bandwidth. Writing the cache costs one prefill, so some vendors charge a small cache-write fee.

Coding agents amplify this. An agent is a loop: the model reads context, issues a tool call, the tool returns a result, the result is put back into context, and the model keeps reading. Every round rereads all previous content. If the context has N tokens and there are T tool calls, total input tokens grow roughly as N times T. As the context grows larger and larger, cost grows quadratically with session length.

Diogo Almeida broke down the bill in the post 'KV Cache Rules Everything Around Me.' Output tokens, which ordinary people think are the most expensive, actually account for only 9% to 18% of the total. The rest, the bulk, is the model repeatedly reading its own history. Martin Alderson simulated a 100-turn, 60,000-token-context session, and cache reads accounted for about 76% of an Opus 5 bill.

A slight change in cache hit rate makes a big difference in the bill. Claude Code's published hit rate is 89%; for the same workload, cost goes from $1.07 million to $1.97 million, an increase of 84%. One team measured 95%; going from 95% to 98% saved about 17%.

So a $200 annual fee producing a $100,000 API-equivalent bill looks like crazy subsidization. SemiAnalysis calculated that a $200/month Claude plan can produce at most $8,000/month of API-equivalent tokens, while OpenAI's plan can reach up to $14,000/month, equivalent to 40x to 70x subsidy at list prices. But this is about API list prices, not the provider's real cost.

One commenter used a Starbucks analogy: giving away a latte with a $6 list price doesn't mean losing $6; the beans and milk might cost only $0.8. The numbers on the API price list and the actual cost of serving are not the same thing.

Shantanu Goel added another side: many vendors have already passed some of the cache benefits on to users. With open-source models, cache reads can be almost free. Anthropic's cache-read price is about 20x cheaper, while it still charges a cache-write fee. He believes inference itself is profitable; KV cache and speculative decoding are more about increasing capacity than generating profit on their own.

Another counterintuitive point is routing. Many people want to use a cheap model for simple steps and an expensive model for hard steps to save output tokens. But output tokens are already the small part. As soon as the second model touches the context, the KV cache is not shared, and the cache must be rebuilt. The output fees saved are easily eaten by cache costs. The same applies to subagents, unless they don't need the full parent context.

If agents run at scale, the significance of self-hosting open-weight models becomes clear: keep the cheapness of cache for yourself instead of handing it to API vendors. At least that's how it is now. Cache-read prices are still falling; Fable 5.1 cut the price from $1 per million tokens to $0.25, showing there is still room to fall.

One commenter added a calculation chart that breaks down cache-read and output costs in more detail.

网友补充的缓存成本计算图

The core of this is not the total number of tokens, but how many times the context is read, and who pays each time it is read. People building agents can watch two numbers: cache hit rate and number of tool calls. The number of tool calls affects the bill more than the amount of tokens in a single call. Vendors don't really want you to know this accounting, because once it is explained clearly, the API price list no longer looks so much like a cost sheet.

发布时间: 2026-11-21 07:59