Claude Fable 5.2: $16.00 on the Blend, the Dearest Rate Here

WhatsApp Channel Join Now

Sixteen dollars per million is what this card costs at a three-to-one input-to-output mix, and the next-highest rate in the set is four. OrcaRouter publishes the two rows the blend comes from. They are on Claude Fable 5.2 pricing: $8.00 per million in and $40.00 out.

The blend, ranked

card

in $/M

out $/M

blend

out ÷ in

Claude Fable 5.2

8.00

40.00

$16.0000

5.0×

Gemini 4 Carbon

2.00

10.00

$4.0000

5.0×

Kimi K3.1 Pro

2.00

10.00

$4.0000

5.0×

GPT-6.1 Sol

2.00

10.00

$4.0000

5.0×

Hy4

0.834

2.501

$1.2508

3.0×

DeepSeek V4.1 Flash 1017

0.15

0.60

$0.2625

4.0×

MiniMax M3.1 Flash

0.15

0.60

$0.2625

4.0×

Hy4 Flash

0.0175

0.0702

$0.0307

4.0×

Sixteen dollars is 4× the next row and 521× the cheapest.

The last column is worth a look because it is nearly constant. Six of eight cards price output at four or five times input, so the blend is a stable multiple of the input price rather than an independent figure.

That stability means the blend column is close to a rescaling of the input column, and a buyer can move between them by remembering that a 3:1 mix puts the blend at four to five times the input rate. The one card that breaks the pattern is Hy4 at 3.0×, on a blend of $1.2508 from an input of $0.834.

Where sixteen dollars lands in absolute terms

A price this high is easier to read as a per-job figure than as a rate.

job

tokens

cost

a 50,000-token answer, short input

60,000

$0.65

a full-window read, cold

1,000,000

$8.0000

a full-window read, warm

1,000,000

$0.2000

a 128,000-token output

128,000

$5.1200

a full-window job, cold

1,128,000

$13.1200

the same job, warm

1,128,000

$5.3200

Thirteen dollars and twelve cents for a full-window job is the figure to carry, because it is the ceiling of the card’s published capability.

The cache row is what makes that number tractable. At $0.20 per million cached, a repeated prefix costs 2.5% of a cold read, so the same job on a warm prefix is $5.32 rather than $13.12 — a 59% saving — the second-steepest cache discount in the set, behind only DeepSeek V4.1 Flash.

So the card’s economics are dominated by how much of each prompt is new. A pipeline re-reading a large fixed document pays a fifth of what a pipeline re-reading nothing pays.

The two thirds that is not money

The price is the headline and it is not the whole published picture.

field

value

position in the set

blend

$16.0000

1st

throughput

147.0 tok/s

2nd

p50

6,153 ms

6th

error rate

14.61 %

3rd highest

ceiling

128,000

5th

window

1,000,000

4th

modalities

text, image, file

5th

benchmark rows

0

joint-last

Second-fastest generation, a mid-table median, and the third-highest error rate. So the premium price does not buy the fastest serving or the cleanest one.

The empty benchmark row is the sharpest entry. This is the most expensive card in the set and it publishes no evaluated score, while five of the eight publish between five and eight rows each.

A buyer paying sixteen dollars a million is paying for something the card does not quantify. That is a legitimate statement about the published fields and not a claim about the model.

The daily series

date

p50 ms

p95 ms

2026-11-14

6,181

10,000

2026-11-15

6,615

10,000

2026-11-16

6,800

10,000

2026-11-17

4,389

10,000

2026-11-18

4,452

10,000

2026-11-19

6,130

10,000

2026-11-20

8,661

10,000

Seven medians between 4,389 and 8,661, a 1.97× spread, and seven capped tails.

The median moved by a factor of two inside the week and the tail never produced a reading. So the card’s published latency cannot be planned against at either end: the centre is unstable and the top is censored.

At a 6,153 ms weekly median the card is not interactive. At 147 tokens a second it is fast once it starts. A caller waits six seconds and then receives text at a rate faster than every card here except Hy4 Flash.

That combination is consistent with a model that spends time before emitting, and it is the same shape K3.1 Pro shows on the cyber tier — fast generation behind a slow first token.

A final note on how to compare a sixteen-dollar blend to anything else here. At a 3:1 input-to-output mix the blend is four to five times the input price on most cards, so the input row alone tells a buyer most of what they need. This card’s input of $8.00 is 8× the band cards and a fifth of nothing else — it is the top of the range. The output row at $40.00 is what makes the blend the highest here, and it is where a workload with long answers feels the price first. A job producing 100,000 output tokens costs $4.00 on output alone, before any input.

One consequence for a buyer weighing this card against its sibling. Claude Fable 5.1 sits on an identical $10.00/$50.00 row and publishes ten rows against this card’s none, with an error rate of 0.36% against 14.61%. Two cards, one price, and every field the predecessor publishes is either absent or better on this one. The published case for choosing this card over its own predecessor is therefore not in the fields at all, and a buyer reading only the card would have no reason to.

A final arithmetic check on the ceiling. A 128,000-token answer at $40.00 per million costs $5.12, and a full-window read adds $8.00 cold or $0.20 warm. So the write is 39% of a cold full-window job and 96% of a warm one. On a card whose cache discount is this steep, the output price is the whole of the bill for any repeated-prefix workload, which makes the $40.00 row the one to plan against.

Against $8.00 per million on the predecessor’s input row, a full-window read costs four times as much for the same million tokens, which is the whole of the price difference in one line.

The card’s five shared rows with its own predecessor are worth one closing note. On GPQA, HLE, long-context recall, SciCode and terminal work, the predecessor publishes all five and scores higher on four. So the newer card’s only published advantage over the card a buyer may already run is first-token latency, and that is a field the predecessor also publishes. A migration decision can be made entirely from the two cards side by side, which is not true of most upgrades here.

Claude Fable 5.2 blends to $16.00 per million tokens from an $8.00 input row and a $40.00 output row — four times the next-highest rate here and 521 times the cheapest — with a cache rate at 2.5% of input.

A full-window job costs $13.12 cold or $5.32 warm. The card generates second-fastest at 147 tokens a second behind a 6,153 ms median, carries a 14.61% error rate, caps its tail on all seven days and publishes no benchmark rows at the highest price in the set.

Sourcing note: the price rows, cache rate, throughput, daily and weekly latency percentiles, error rate, window, ceiling, modality list and token volume are read from the Claude Fable 5.2 model API and card page on orcarouter.ai on 2026-11-20, with latency, throughput and error rates being OrcaRouter production telemetry for 2026-11-14 to 2026-11-20. The blends, full-window costs, cache savings and rankings are our own arithmetic at a 3:1 input-to-output mix. A reading of exactly 10,000 ms is the reporting cap, not a measurement. No competitor product or benchmark site is named. Checked 2026-11-20.

Similar Posts