GitHub
Alexandre Gensse
Data analysis · Claude cost vs quality

Claude models head-to-head Fable vs Opus vs Sonnet vs Haiku

Pick two Claude models and compare them effort by effort, on the same scales as the main comparison. Opus 5.5 at medium gives the same quality as Fable 5.1 at max for 8.3× less per task; Opus 5.5 at high gives the same quality as Sonnet 5.5 at max for 2.6× less per task.

Updated
21 Oct 2026

Compare two Claude models

Fable
Opus
Sonnet
Haiku

Fable 5.1vsOpus 5.5

20×50×100×200×64 %66 %68 %70 %72 %74 %76 %cost per task, × the cheapest couple (log scale)expected score on the panellowmediumhighxHighmaxlowmediumhighxHighmax

Opus 5.5 at Medium gives the same quality as Fable 5.1 at Max for 8.3× less per task.

By effort level cost per task × the cheapest couple · expected score on the benchmark panel

EffortFable 5.1Opus 5.5
costscorecostscore
low70×64.8 %25×66.0 %
medium84×67.2 %35×70.8 %
high102×68.2 %43×72.2 %
xHigh186×69.8 %74×73.9 %
max293×70.6 %194×74.7 %
Every pair, written out

Fable 5.1vsOpus 5.5

20×50×100×200×64 %66 %68 %70 %72 %74 %76 %cost per task, × the cheapest couple (log scale)expected score on the panellowmediumhighxHighmaxlowmediumhighxHighmax

Opus 5.5 at Medium gives the same quality as Fable 5.1 at Max for 8.3× less per task.

By effort level cost per task × the cheapest couple · expected score on the benchmark panel

EffortFable 5.1Opus 5.5
costscorecostscore
low70×64.8 %25×66.0 %
medium84×67.2 %35×70.8 %
high102×68.2 %43×72.2 %
xHigh186×69.8 %74×73.9 %
max293×70.6 %194×74.7 %

Fable 5.1vsSonnet 5.5

10×20×50×100×200×60 %65 %70 %cost per task, × the cheapest couple (log scale)expected score on the panellowmediumhighxHighmaxlowmediumhighxHighmax

Sonnet 5.5 at xHigh gives the same quality as Fable 5.1 at Max for 8.9× less per task.

By effort level cost per task × the cheapest couple · expected score on the benchmark panel

EffortFable 5.1Sonnet 5.5
costscorecostscore
low70×64.8 %12×58.5 %
medium84×67.2 %13×61.7 %
high102×68.2 %18×66.9 %
xHigh186×69.8 %33×70.9 %
max293×70.6 %115×72.8 %

Fable 5.1vsHaiku 5.5

1×2×5×10×20×50×100×200×50 %55 %60 %65 %70 %cost per task, × the cheapest couple (log scale)expected score on the panellowmediumhighxHighmaxlowmediumhighxHighmax

By effort level cost per task × the cheapest couple · expected score on the benchmark panel

EffortFable 5.1Haiku 5.5
costscorecostscore
low70×64.8 %1.0×51.0 %
medium84×67.2 %1.4×55.5 %
high102×68.2 %2.2×58.6 %
xHigh186×69.8 %4.3×61.8 %
max293×70.6 %12×64.2 %

Opus 5.5vsSonnet 5.5

10×20×50×100×200×60 %65 %70 %75 %cost per task, × the cheapest couple (log scale)expected score on the panellowmediumhighxHighmaxlowmediumhighxHighmax

Opus 5.5 at High gives the same quality as Sonnet 5.5 at Max for 2.6× less per task.

By effort level cost per task × the cheapest couple · expected score on the benchmark panel

EffortOpus 5.5Sonnet 5.5
costscorecostscore
low25×66.0 %12×58.5 %
medium35×70.8 %13×61.7 %
high43×72.2 %18×66.9 %
xHigh74×73.9 %33×70.9 %
max194×74.7 %115×72.8 %

Opus 5.5vsHaiku 5.5

1×2×5×10×20×50×100×200×50 %55 %60 %65 %70 %75 %cost per task, × the cheapest couple (log scale)expected score on the panellowmediumhighxHighmaxlowmediumhighxHighmax

By effort level cost per task × the cheapest couple · expected score on the benchmark panel

EffortOpus 5.5Haiku 5.5
costscorecostscore
low25×66.0 %1.0×51.0 %
medium35×70.8 %1.4×55.5 %
high43×72.2 %2.2×58.6 %
xHigh74×73.9 %4.3×61.8 %
max194×74.7 %12×64.2 %

Sonnet 5.5vsHaiku 5.5

1×2×5×10×20×50×100×50 %55 %60 %65 %70 %75 %cost per task, × the cheapest couple (log scale)expected score on the panellowmediumhighxHighmaxlowmediumhighxHighmax

By effort level cost per task × the cheapest couple · expected score on the benchmark panel

EffortSonnet 5.5Haiku 5.5
costscorecostscore
low12×58.5 %1.0×51.0 %
medium13×61.7 %1.4×55.5 %
high18×66.9 %2.2×58.6 %
xHigh33×70.9 %4.3×61.8 %
max115×72.8 %12×64.2 %

Fable 5.1vsFable 5

100×200×60 %65 %70 %cost per task, × the cheapest couple (log scale)expected score on the panellowmediumhighxHighmaxlowmediumhighxHighmax

Fable 5.1 at Medium gives the same quality as Fable 5 at Max for 3.2× less per task.

By effort level cost per task × the cheapest couple · expected score on the benchmark panel

EffortFable 5.1Fable 5
costscorecostscore
low70×64.8 %69×60.5 %
medium84×67.2 %95×64.0 %
high102×68.2 %115×65.9 %
xHigh186×69.8 %165×66.9 %
max293×70.6 %270×67.8 %

Opus 5.5vsOpus 5

20×50×100×200×65 %70 %75 %cost per task, × the cheapest couple (log scale)expected score on the panellowmediumhighxHighmaxlowmediumhighxHighmax

Opus 5.5 at Medium gives the same quality as Opus 5 at Max for 4.2× less per task.

By effort level cost per task × the cheapest couple · expected score on the benchmark panel

EffortOpus 5.5Opus 5
costscorecostscore
low25×66.0 %39×62.5 %
medium35×70.8 %57×66.1 %
high43×72.2 %84×68.0 %
xHigh74×73.9 %111×69.2 %
max194×74.7 %148×69.6 %

Sonnet 5.5vsSonnet 5

10×20×50×100×50 %55 %60 %65 %70 %75 %cost per task, × the cheapest couple (log scale)expected score on the panellowmediumhighxHighmaxlowmediumhighxHighmax

Sonnet 5.5 at Medium gives the same quality as Sonnet 5 at Max for 8.5× less per task.

By effort level cost per task × the cheapest couple · expected score on the benchmark panel

EffortSonnet 5.5Sonnet 5
costscorecostscore
low12×58.5 %16×48.3 %
medium13×61.7 %23×52.8 %
high18×66.9 %33×55.9 %
xHigh33×70.9 %50×57.5 %
max115×72.8 %109×60.1 %

Sonnet 5's cost is size-sensitive: a verbose model, its cost swings widely between short and long agentic tasks, hence its wide interval.

How to read these comparisons

The values are the ones the main page shows, fitted from public measurements taken on the same tasks. One model is cheaper or higher than the other at an effort level when the fitted difference puts it ahead with at least 84 % probability, the level of the intervals shown everywhere on the site; otherwise the two are level within the uncertainty. A match is the cheapest setting of the other model that the first does not out-score at that level. Costs are what a whole task cost, as each source measured it (the run's actual spend, cache included), not the price per token. So the cost already counts how many tokens each model uses: a model with a lower price per token can still cost more per task if it writes many more tokens.

Alexandre Gensse · github.com/alexandregensse-blip · Claude models compared: cost vs quality at every effort level

All figures are indicative and for informational purposes only — derived from public third-party measurements, not an official benchmark, and not affiliated with or endorsed by Anthropic. Verify before relying on any value.

Data and text under CC BY 4.0 · code under MIT · download the data (CSV)