Anthropic's Claude Fable 5.1 pricing announcement is a masterclass in reading a price sheet carefully. The headline number didn't change: $10 per million input tokens, $50 per million output tokens, identical to Fable 5. The change that actually matters is buried a row below: cache reads dropped 75%, from $1.00 to $0.25 per million tokens.
Whether that row saves you 45% or exactly 0% depends entirely on how your prompts are structured. Let's do the math.
The official table
| Rate | Price (per 1M tokens) | Notes |
|---|---|---|
| Input | $10 | unchanged from Fable 5 |
| Output | $50 | unchanged from Fable 5 |
| Cache read | $0.25 | down 75% from $1.00 — the headline change |
| Cache write | ~$12.50 | confirm current on Anthropic's pricing page |
| Context window | 1M tokens | 128K max output |
Anthropic's own estimates for the total effect: ~25% cheaper than Fable 5 for typical workloads, up to ~45% for heavily agentic work — and per third-party analysis, almost all of that comes from the cache-read cut alone.
Why caching is 40x cheaper than fresh input
In any agent session, the same large prefix gets re-sent on every step: system prompt, tool definitions, codebase context. Once cached, that input bills at $0.25 instead of $10 — a 40x discount on the cached portion of your input.
The corollary is the part the launch posts skip: a workload with a low cache hit-rate gets none of this discount. Fresh $10/$50 tokens is Opus-tier money, and it's what you'll pay if your prompts churn.
Two worked examples — one win, one trap
Example 1: a coding agent loop (the big winner). 40 steps, each re-reading a 50K-token cached prefix, 2K output per step:
| Item | Fable 5 | Fable 5.1 |
|---|---|---|
| 2M cached input (40 × 50K) | $2.00 | $0.50 |
| 80K output | $4.00 | $4.00 |
| Session total | $6.00 | $4.50 (−25%) |
Example 2: document Q&A with fresh input. 10 questions against a 100K-token document, no caching:
| Item | Fable 5 | Fable 5.1 |
|---|---|---|
| 1M fresh input | $10.00 | $10.00 |
| 20K output | $1.00 | $1.00 |
| Total | $11.00 | $11.00 (0%) |
Example 2 is the trap. Without caching, 5.1 costs exactly what 5 did. The discount isn't a property of the model — it's a property of your prompt structure. Stable parts first, never touched; variable parts trailing.
Monthly numbers, with the caveat attached
For a solo developer running a coding agent ~4 hours/day (roughly 600 agent sessions/month, cached-prefix profile), arithmetic on the official rates gives:
| Profile | Fable 5 / month | Fable 5.1 / month |
|---|---|---|
| Light (10 sessions/day, short contexts) | ~$15–30 | ~$11–22 |
| Heavy agentic (long contexts, many steps) | ~$150–300 | ~$85–170 |
| Fresh-input heavy (no caching) | ~$100–200 | ~$100–200 (no change) |
These are arithmetic, not measured bills. Your prompt structure moves them more than any rate does.
Four ways to actually pay less
- Cache everything cacheable — stable system prompts and tool definitions go first; that's where the 75% cut lives.
- Right-size the effort level — adaptive reasoning is always on in 5.1, and high effort on routine tasks burns $50/M output tokens for nothing.
- Route by task — Fable 5.1 for the hard 20%, cheaper models for the rest. The flagship-vs-flagship math (5.1's $0.25 cache read vs GPT-6 Astra's $1.00) is in Fable 5.1 vs GPT-6 Astra.
- Pay-as-you-go instead of subscription — a gateway like TeamoRouter passes through official rates with tiered usage discounts, no subscription needed, Alipay/WeChat supported. One subtlety: gateways also unify caching behavior across providers, so a workload that bounces between models keeps its cache discounts instead of re-paying fresh-input rates at each vendor.
Audit us — and anyone else selling you access
Since TeamoRouter resells Fable 5.1, don't take the "cheaper via gateway" line on faith — from us or anyone:
- Log the
usagefield from every response — input, output, and cache_read_input_tokens — and compute your own effective $/M. - Run a fixed 100-request benchmark suite through the official API and the gateway on the same day; rates change, trust your logs.
- Check whether the gateway passes through Anthropic's cache pricing or flattens it — a flattened rate can quietly destroy the 75% discount.
Subscription note for completeness: Fable 5.1 is included in Claude Max plans and premium Team/Enterprise seats, with up to half your weekly usage limit allocatable to it — a fine route if your volume fits a plan. Pay-as-you-go wins when it doesn't.
CTA: TeamoRouter is a multi-model API gateway — Claude Fable 5.1 and GPT-6 Astra are both available at official rates behind one key and one base URL, with per-model cost dashboards so the cache math above is visible in your own logs. Pay-as-you-go, Alipay/WeChat supported. More cost breakdowns on the TeamoRouter blog.
Top comments (0)