VynarisEarly betaGet your API key

Blog · 2026-09-02 · Vynaris Team

Claude Fable 5.1 cache reads cost $0.25/MTok: why agent bills fall 25%-45%

Claude Fable 5.1 cuts cache reads 75%. A disclosed agent session falls from $8.50 to $5.50, but three breaking changes can erase the saving.

Claude Fable 5.1 cuts cache reads from $1 to $0.25 per million tokens while holding base input and output at $10/$50. On our disclosed cache-heavy agent session, the bill falls from $8.50 to $5.50, a 35.3% cut. Anthropic reports about 25% for typical workloads and up to about 45% for highly agentic work. Prices verified 2026-09-02.

TL;DR

Verdict table

Decision receipt                       Claude Fable 5  Claude Fable 5.1  Change
-------------------------------------  --------------  ----------------  -----------------
Base input per 1M tokens               $10.00          $10.00            None
Cache read per 1M tokens               $1.00           $0.25             -75%
Five-minute cache write per 1M tokens  $12.50          $12.50            None
One-hour cache write per 1M tokens     $20.00          $20.00            None
Output per 1M tokens                   $50.00          $50.00            None
Batch input/output per 1M tokens       $5.00/$25.00    $5.00/$25.00      None
Disclosed agent session                $8.50           $5.50             -$3.00, or -35.3%

The migration is a clear price win only for workloads that already produce cache hits. A new model name does not make changing prefixes cacheable.

The bill moves in proportion to the old cache-read share

Let s be cache reads as a share of the old bill. The cache-read rate falls 75%, while the other listed rates stay fixed. The price-only saving is therefore:

total bill saving = 75% × s

If cache reads were 10% of the old bill, the total falls 7.5%. At a 33.3% share, it falls 25%. At a 60% share, it falls 45%.

Target total saving  Required cache-read share of old bill
-------------------  -------------------------------------
5%                   6.7%
10%                  13.3%
25%                  33.3%
45%                  60.0%
Line chart showing total bill savings as cache reads occupy more of the old bill
A 75% cache-read cut produces a 25% total saving at a 33.3% old-bill cache share and a 45% saving at 60%. Source: Anthropic list prices; verified 2026-09-02.

This threshold uses bill share, not token share. One million output tokens costs $50, while one million old cache reads cost $1. Equal token counts do not carry equal weight.

A cache-heavy agent session falls from $8.50 to $5.50

Consider one agent session with 4 million cache-read tokens, 200,000 fresh input tokens, and 50,000 output tokens. We exclude writes, tools, retries, regional premiums, and partner-platform markups so the changed meter stays visible.

Claude Fable 5 costs:

4M × $1/M + 0.2M × $10/M + 0.05M × $50/M = $4 + $2 + $2.50 = $8.50

Claude Fable 5.1 costs:

4M × $0.25/M + 0.2M × $10/M + 0.05M × $50/M = $1 + $2 + $2.50 = $5.50

The absolute saving is $3 per session. Across 1,000 identical sessions, the bill moves from $8,500 to $5,500. Put measured cache-read, fresh-input, and output counts into the LLM cost calculator before forecasting a production saving.

Our prompt-caching break-even analysis covers the write side. The five-minute write still costs 1.25 times base input. The one-hour write still costs 2 times. A cheaper read does not rescue a prefix that changes before reuse.

What Anthropic's 25%-45% range does and does not prove

Anthropic reports an indexed cost move from 100 to about 75 for its typical workload and from 100 to about 55 for highly agentic work. The launch-page caption says these are default-effort workloads measured over four weeks of actual August 2026 usage. Typical usage spans Claude Enterprise, Claude Code, and the API. The agentic group is context-heavy and tool-heavy.

Those are public production aggregates, not a published token trace. The chart does not expose fresh input, cache writes, cache reads, output, tools, or retries as token counts. We can reproduce the 25% and 45% index reductions. We cannot honestly convert them into one universal tokens-per-task recipe.

Under a fixed-token price model, those reductions imply old-bill cache-read shares of 33.3% and 60%. Model behavior can move the observed result. Fable 5.1 may take a different number of turns or tokens than Fable 5. Our full-conversation token-tax analysis shows why one extra agent turn can re-bill a large history.

Three migration checks before claiming the saving

Anthropic's migration guide names three breaking changes for Messages API integrations.

  1. Forced tool_choice values any and named tool return a 400. Use auto with an explicit instruction and strict schema, or use JSON outputs.
  2. Fable 5.1 can read thinking blocks from earlier Claude models, but most older models cannot read Fable 5.1 blocks. A fallback can drop them and re-plan, adding cost and latency.
  3. Editing, reordering, or removing earlier turns can invalidate Fable 5.1 thinking blocks. Newer accounts receive a 400 by default. Keep history append-only, strip invalid blocks, or use the documented drop control.

The guide also warns that long loops may issue fewer parallel tool calls. More turns mean more context reads. Measure the full cost per task, not just the cheaper line item.

What this means for routing

Read the usage object before changing a model routing rule. Compute cache-read dollars as a share of the current Fable 5 bill. Multiply that share by 75% for a first price-only estimate.

Then run the same accepted-task eval on both versions. Track cache reads, fresh input, output, attempts, tool calls, latency, and fallback behavior. Keep batch processing separate: Anthropic still lists a 50% discount on base input and output, while cache multipliers stack with other pricing modifiers.

Honest tradeoff: a cheaper cache read can hide a worse loop

Do not migrate a stable revenue path only for a forecasted 5% token saving. A forced-tool failure, a dropped thinking block, or one extra serial tool turn can erase it. The price cut is strongest for context-heavy loops with proven cache hits and migration tests.

The opposite mistake is dismissing the change because $10/$50 stayed fixed. At a 60% old-bill cache share, the price-only saving is 45%. That is large enough to change unit economics without changing the model's headline rate.

Sources