智能体
dev.to ·
@szp2005
Fable 5.1 的缓存读取降价：何时才比 Opus 5 更便宜
文章分析 Claude Fable 5.1 的定价变化：输入 $10、输出 $50 每百万 token 与 Fable 5 相同，但缓存读取价从 $1 降至 $0.25，而 Opus 5 为 $5/$25、缓存读取 $0.50。作者给出交叉点公式 R = 20U + 100O（R 为每请求缓存 token，U 为新输入，O 为输出），代入 agent 场景（100K 缓存、2K 新输入、1K 输出）后盈亏平衡点约为 140K 缓存 token，此时两者均约 $0.105/请求；超过后 Fable 5.1 更便宜，200K 缓存时约省 11%。作者同时指出首次分析遗漏缓存写入成本：若按 1.25x 输入价计，写入 200K 前缀在 Fable 5.1 上多花 $1.25，而每请求仅省 $0.015，需复用同一前缀 80 次以上才能回本，前缀频繁过期或反复重写上下文时可能无法达到交叉点。建议从日志中取缓存 token、新输入和输出三项均值，并统计前缀复用次数后再决定是否迁移。
AI推荐理由 文章对 Claude Fable 5.1 与 Opus 5 的定价做了可验证的量化拆解：Fable 5.1 缓存读取价从 $1 降至 $0.25/百万 token，输入 $10、输出 $50 不变；给出交叉点公式 R = 20U + 100O，并代入 agent 场景（100K 缓存、2K 新输入、1K 输出）得出约 140K 缓存 token 的盈亏平衡点。作者还主动指出首次分析遗漏的缓存写入成本（按 1.25x 输入价估算），提示前缀频繁过期时可能永远无法回本，属于有具体数字和局限说明的分析。局限是数据来自作者自建价格表，非官方定价页，且 Fable 5.1 与 Opus 5 的命名和发布时间无法在输入中独立核实，需以官方价格页为准。
72
正文
Claude Fable 5.1 shipped on September 1 with the same $10 input / $50 output per million tokens as Fable 5. One number moved: cache reads dropped from $1 to $0.25 per million.
On the price sheet it still costs twice what Opus 5 does ($5 / $25). But look at the cache-read column and the order flips. Fable 5.1 charges $0.25 there, Opus 5 charges $0.50. So which one is cheaper depends on how much of each request is served from cache, and I wanted the actual crossover instead of a vibe.
A request bill has three parts
Cached tokens times the cache-read price, fresh input times the input price, output times the output price. Fable 5.1 only changed the first term, so the saving over Fable 5 is exactly as big as that term's share of your bill.
Two workloads I priced out from the table on my site:
Per request
Fable 5
Fable 5.1
Opus 5
Agent: 100K cached, 2K new input, 1K output
$0.170
$0.095
$0.085
Chat: 5K cached, 1K new input, 800 output
$0.055
$0.051
$0.028
In the agent loop, cache reads are about 60% of the Fable 5 bill, so the price cut takes 44% off each request. That lines up with the "around 45% for heavy agent use" figure in the launch coverage.
The chat case barely moves. Output dominates, and Opus 5 stays roughly half the price. Switching there is just paying more.
The crossover
Fable 5.1 pays double for fresh input and output, and half for cache reads. Set the two bills equal:
0.25 * R = 5 * U + 25 * O
R = 20 * U + 100 * O
Enter fullscreen mode
Exit fullscreen mode
R is cached tokens per request, U is fresh input, O is output. Plug in the agent numbers above and the line sits at 140K cached tokens. Both models cost $0.105 a request there.
Past that point the gap keeps widening. At 200K cached, Fable 5.1 is about 11% cheaper. Double the cache again and it's more than a quarter cheaper.
Output length is what pushes the line out. Every extra thousand output tokens needs another hundred thousand cached tokens before Fable 5.1 catches up.
The part I got wrong the first time
My first pass stopped there and the conclusion was "long-context agents should move to Fable 5.1." Then I remembered that tokens have to get into the cache before they can be read from it.
Anthropic bills a 5-minute cache write at 1.25x the input price for Opus 5. My price table has no separate write price for Fable 5.1. If it follows the same 1.25x rule, writing a 200K prefix costs $1.25 more on Fable 5.1 than on Opus 5. At 200K cached, Fable 5.1 saves $0.015 per request. That's more than 80 requests on the same prefix before the write premium is paid back.
So if your prefix expires every five minutes, or your agent keeps rewriting its context, you may never reach the crossover at all. Measure how many requests actually reuse each prefix before you switch.
How I'd decide
Pull three averages from your logs: cached tokens, fresh input and output per request. Then check how many requests reuse the same prefix.
Below 20U + 100O cached tokens: stay on Opus 5.
Above it, but prefixes get rebuilt every few dozen requests: the write premium eats the difference. Stay on Opus 5.
Above it, with one long-lived prefix (a coding agent in a long session, Q&A over one big fixed document): Fable 5.1 wins per request, and it scores 66 vs 63 on the quality index I track.
Already on Fable 5: switch. Nothing got more expensive.
I built llmabacus.com, the LLM API price table these numbers come from. The full write-up, in Chinese, is at https://www.llmabacus.com/articles/fable-5-1-cache-read-vs-opus-5