Between Claude Fable 5 and Claude Fable 5.1, three of the four rates did not move. Input stayed at $10.00 per million, output at $50.00, and the cache write at $12.50 [OURS]. The fourth moved a long way: the cached read fell from $1.00 to $0.25, a quarter of what it was. The claude fable 5.1 price query is answered by that row. The card is served by OrcaRouter.
The four lines, before and after
| Line | Fable 5 | Fable 5.1 |
| Input | $10.00 | $10.00 |
| Output | $50.00 | $50.00 |
| Cache write | $12.50 | $12.50 |
| Cached read | $1.00 (10%) | $0.25 (2.5%) |
A 2.5 percent cache read is the cheapest in this family. Claude Opus 4.6 and Claude Sonnet 5 both read cache at 10 percent of input; the older Fable 5 did too. Fable 5.1 is the outlier at a fortieth of the input price.
Nothing else in the shape changed. The write still costs 1.25 times the input it replaces, so the premium is still $2.50 per million tokens, and the question is still whether a prefix gets reused.
What a 2.5 percent read does to the break-even
The break-even formula has not changed, only its inputs.
Writing one million tokens costs $12.50 instead of $10.00 — a premium of $2.50. Reading that million back costs $0.25 instead of $10.00 — a saving of $9.75 per reuse. The premium is repaid in 0.26 reuses, which in practice means the first time the prefix is read.
Take a 500,000-token instruction block and forty calls against it in a day:
| Per call | Forty calls | |
| No cache | $5.00 | $200.00 |
| Cache once, read 40 times | $6.25 + 40 × $0.125 | $11.25 |
The cached path is 94.4 percent cheaper [OURS arithmetic on the published rates]. On the older card, the same workload costs $5.00 × 40 against $6.25 + 40 × $0.50, so $200 against $26.25 — still a large saving, but 2.3 times what the new cache line charges.
That is the whole point of the change: for a workload built on a fixed prefix, the read cost has stopped being a rounding error and become a rounding error’s rounding error.
The number that argues against it
The same seven-day window that makes the cache line attractive reports an error rate of 41.43 percent [OURS]. That is the highest of any card in this batch that served real traffic — 14,518,248 tokens over the window, so the sample is not thin.
An error rate that high does two things to the arithmetic above.
It doubles the cache economics in the wrong direction. A cached prefix still has to be written, and a failed request does not refund the write. If two in five calls fail and are retried, the effective write cost per successful call rises by 1.67 times while the read saving per successful call stays the same.
It caps the workload this card is right for. A batch job can retry and absorb a 41 percent failure rate at the cost of latency. An interactive product cannot: two failures in five turns is not a reliability problem, it is a product problem.
The honest reading is that the cache line is a genuine improvement priced by someone who expects heavy, patient reuse — and the telemetry says the route is not yet steady.
Latency, which is fine
Median time to first token is 3,586 ms and p95 is 6,666 ms [OURS]. Unlike several cards in this batch, the p95 is a measured value rather than the 10,000 ms recording ceiling, so the distribution is real: the slow tail is 6.7 seconds, not unmeasured.
Output speed is 83.5 tokens per second. For a 1,000-token answer that is twelve seconds after the first token, which is ordinary for a frontier-priced model and slower than the cheap tier by a factor of three or so.
Taken together, the latency is the least interesting part of this card. It behaves like a model of its class; the cache line and the error rate are the two numbers that are unusual.
What to do with a card like this
Three rules.
Price the write, not the read. At $0.25 the read is close to free and no longer decides anything. What decides the bill is how many times you write, and a 41 percent error rate makes writes more frequent than the call count suggests.
Retry in the routing layer, not the application. A failed call is cheap at the token level and expensive at the cache level. Routing that retries before the write is re-issued keeps the cache economics from degrading.
Treat a falling cache price as a statement of intent. A vendor cutting the read price by four times is telling you the intended workload is long-prefix reuse. Whether the route is ready for that intent is a separate question, and this card’s error rate says not yet.
The takeaway
Claude Fable 5.1 keeps $10.00 input, $50.00 output and a $12.50 cache write, and cuts the cached read to $0.25 per million tokens — 2.5 percent of input, four times cheaper than Fable 5’s $1.00 [OURS].
For a prefix reused even once, that is the cheapest reading in the family. For a route reporting a 41.43 percent error rate over seven days, the saving is real and the reliability is the thing to fix first.
Sourcing note: the four rates for both generations, the context window, the output ceiling, the nineteen and ten benchmark rows, and the seven-day p50, p95, error rate, output speed and token total are read from the OrcaRouter public model cards for `anthropic/claude-fable-5.1` and `anthropic/claude-fable-5` on 2026-10-23 [OURS]. The 2.5 percent and 1.25× ratios, the break-even and the per-day tables are our own arithmetic on the published rates. The impression figure is Search Console for the 28 days to 2026-10-21 [OURS]. No competing gateway or inference marketplace is named. All values checked 2026-10-23. Click for more details.
