Claude Opus 5.5 vs Claude Fable 5.1: The Cheaper Model Won the Independent Index

Ant​hropic now sells two models at the top of its lineup, and the one that costs two and a half times more is not the one at the top of the table. Claude Opus 5.5, released September 22, 2026 at $4 per million input and $20 per million output tokens, scores 58 on the Artificial Analysis Intelligence Index. Claude Fable 5.1, which has held the flagship position since September 1 at $10 in and $50 out, scores 53. Both figures come from the same evaluator, both were measured in the same configuration family, and the gap runs the opposite way to the price tags. That is the fact this comparison has to start from, because almost every launch write-up from the past week frames Opus 5.5 as the discount version of Fable 5.1 — a cheaper model that "matches" the expensive one on most work. The independent number says the discount model is ahead.

Whether that should change what you deploy is a different question, and the answer depends less on the five-point gap than on two settings nobody puts in a headline: which effort level each model defaults to, and which one can be made fast.

The independent numbers, and the one thing they share

Artificial Analysis publishes one configuration per model, and for both of these it is labelled "Adaptive Reasoning, Max Effort, Default Fallback". Same label, same evaluator, same run — which makes this the cleanest head-to-head either model has:

• Intelligence Index — Claude Opus 5.5 58, ranked #1 of 210 vs Claude Fable 5.1 53, ranked #4

• Output speed — Claude Fable 5.1 65.8 tokens/sec, below the 71.0 board median vs Claude Opus 5.5 not yet published on the board

• Time to first token — Claude Fable 5.1 274.43s vs a board median of 3.83s; Claude Opus 5.5 not yet published

• Verbosity on the Index — Claude Opus 5.5 generated 260M output tokens, against a median of 88M across the board

• Price — Claude Opus 5.5 $4.00 / $20.00 per million vs Claude Fable 5.1 $10.00 / $50.00

Two of those rows need reading rather than quoting. The first is the "Max Effort" label: neither model's score reflects its shipping default, and the two defaults are not the same — more on that below. The second is the verbosity row, which cuts against the way Opus 5.5 has been sold. Anthropic's efficiency story is that the model reaches the same answer with fewer tokens, and the launch material cites partners reporting a third to a half fewer output tokens. On the Index, measured rather than quoted, Opus 5.5 emitted 260M tokens against a board median of 88M — roughly three times more talkative than the typical model it is being compared to. Both things can be true: it may use fewer tokens than Claude Opus 5 while still being a verbose model in absolute terms. But if you budgeted from the "fewer tokens" line and not from your own bill, the independent measurement is the one to check.

What the extra $6 per million buys

Strip out the benchmarks and the difference is a rate card. Everything here is from Anthropic's published pricing page, verifiable today:

• Input — $4.00 per million on Opus 5.5 vs $10.00 on Fable 5.1

• Output — $20.00 per million vs $50.00

• Cache read — $0.20 per million on Opus 5.5 vs $0.25 on Fable 5.1

• Cache multiplier — Opus 5.5 bills cache hits at 0.05x base input; Fable 5.1 bills them at 0.025x, a better multiplier on a higher base

• Cache write — $5 per million for a 5-minute window and $8 for an hour on Opus 5.5, vs $12.50 and $20 on Fable 5.1

• Batch API — Opus 5.5 halves to $2.00 / $10.00; Fable 5.1 halves to $5.00 / $25.00

• Fast mode — Opus 5.5 supports it at $8.00 / $40.00 for up to roughly 2.5x the output speed; Fable 5.1 does not support fast mode at all

The cache line is the one to look at twice, because the two models invert the usual pattern. Fable 5.1 has the more aggressive multiplier — 2.5% of base input against Opus 5.5's 5% — but because its base is $10 rather than $4, its cache read is still the more expensive of the two in absolute dollars. There is no configuration in which the premium model wins on a token of cached context. And the multiplier is not something you can port: an agent loop tuned to Fable 5.1's caching behaviour will pay more per cache hit on Opus 5.5 proportionally, even as it pays less per fresh token.

Fast mode is the sharper asymmetry. Anthropic's documentation lists fast mode for Claude Opus 5.5, Claude Opus 5 and Claude Opus 4.8 — Fable 5.1 is absent from that list. So the cheaper model has a latency lever the expensive one simply does not, at $8/$40, which is still below Fable 5.1's standard $10/$50 output rate. If your workload is interactive enough that a slow first token is a product problem, note that Fable 5.1's measured time to first token is 274 seconds. That number is an artifact of max-effort reasoning, not a defect, but it is the number an evaluator recorded.

The defaults are not the same, and that is the trap

Anthropic's own model documentation puts the two models side by side, and the row that decides most real comparisons is not the price. It is the default effort level: medium for Claude Opus 5.5, high for Claude Fable 5.1. Both run adaptive thinking that cannot be turned off; depth is controlled only through the effort parameter, on a scale that runs low, medium, high, xhigh, max.

This is why the launch line "Opus 5.5 matches Fable 5.1 on most work" and the benchmark tables underneath it look like they disagree. Anthropic's head-to-head rows for Opus 5.5 are frequently labelled with a higher effort setting than the default — the Terminal-Bench 4.0 figure of 66.4% against Fable 5.1's 55.8% was run at xhigh effort, not at medium. Meanwhile Fable 5.1's own numbers were produced at its higher default. Two models compared at different effort settings are not being compared at all, and Anthropic says as much in its own launch material when it cautions that benchmark margins at this capability level are a less reliable guide to real-world difference than the scores imply.

Practically, that turns the decision into a tuning exercise rather than a selection. If you move from Fable 5.1 to Opus 5.5 and carry your effort setting across unchanged, you have silently changed how much thinking the model does, and every quality and cost number you produce afterwards is confounded. Anthropic's guidance is to test multiple effort levels rather than port the old model's settings. That is cheap to do and almost nobody does it, because it means running your own evals twice.

The one place the pairing earns its keep

The pattern that justifies keeping both is the advisor loop: Opus 5.5 doing the work, Fable 5.1 consulted on the hard decisions. Anthropic measured it, and it does add accuracy — at higher cost and higher latency, which is the trade you would expect from putting the slower model in the loop. What has changed is the arithmetic behind it. When the capability gap was wider, paying $10/$50 to supervise a $5/$25 worker was obviously worth it. Now that the worker outscores the advisor on the independent index, the advisor's contribution narrows to the specific tasks where its own evals beat Opus 5.5, and those are the tasks you have to identify yourself. The loop still makes sense for a hard subset; it stops making sense as a blanket configuration.

Both are one key away

The migration detail that catches teams out here is not the API surface — it is that these are the same vendor's models with different effort scales, different cache multipliers and different default settings, and comparing them honestly means running both against your own prompts at matched configuration. Claude Opus 5.5 is on OrcaRouter at Anthropic's own $4.00 / $20.00, and Claude Fable 5.1 is on OrcaRouter at $10.00 / $50.00 — provider list price passed through with 0% markup, so both numbers move the day Anthropic moves them and neither carries a second contract or a second SDK.

That is what makes the split practical rather than theoretical. Sending the bulk of your traffic to Opus 5.5 and pinning a routing rule that escalates the genuinely hard cases to Fable 5.1 is a configuration on one endpoint, not two integrations — and composing a call so one model drafts while the other verifies is a routing-DSL line rather than a pipeline you have to build and maintain. If the escalation path turns out to be wrong for your workload, automatic failover keeps the request landing on a model you have already characterised instead of failing outright.

What would settle it

The honest state of this comparison is that Anthropic has published one table where the cheaper model wins most rows, Artificial Analysis has published another where it wins outright, and both were run at effort settings you do not ship. Neither is a controlled head-to-head at matched defaults, and no third party has run one. The gap that survives all of that — that Claude Opus 5.5 is materially cheaper on every line of the rate card, supports a fast mode Fable 5.1 does not, and is not behind on the only independent score either model has — is large enough that the burden of proof has moved. If you are on Fable 5.1 by default, the question is no longer whether Opus 5.5 is good enough; it is which specific tasks on your workload fail at medium effort, because those are the only ones you are paying the premium for.