Fable 5.1 vs Opus 5: Twice the Price, Behind on Every Published Benchmark

Claude Fable 5.1 costs twice what Claude Opus 5 costs. On the benchmarks that exist, Opus 5 scores higher. And Anthropic's own documentation tells you to start with Opus 5 and reach for Fable only when your evals fall short. That combination makes this an unusually easy comparison to get right, and an unusually easy one to get wrong if you assume the pricier model is the better one.

Fable 5.1 is $10 per million input tokens and $50 per million output. Opus 5 is $5 and $25. The ratio is exactly two to one in both directions, and Fable 5.1's per-token price is unchanged from Fable 5 — the widely quoted “25% cheaper” figure is a whole-workload saving that comes entirely from cache reads, not a cut to the rate.

Anthropic published no benchmark scores for Fable 5.1 at launch, so the comparison has to run against the published Fable 5 baseline. On that basis Opus 5 leads on all three public measures:

• ▸<strong>SWE-bench Verified:</strong> Opus 5 at 96.0% against Fable 5's 95.0%.
• ▸<strong>Frontier-Bench v0.1:</strong> Opus 5 at 43.3% against Fable 5's 33.7%, which is the widest of the three gaps.
• ▸<strong>GDPval-AA v2 knowledge work:</strong> Opus 5 at 1,861 Elo against Fable 5's 1,747.

Anthropic states that Fable 5.1 improves on Fable 5, with the gap widest at higher effort levels, so treat those Fable numbers as a floor rather than a verdict. But a floor is what we have, and it sits below Opus 5 on every published test while costing twice as much.

This is the part most comparisons skip. Anthropic's own model documentation says to start with Claude Opus 5 for most workloads, and to use Fable 5.1 for demanding reasoning and long-horizon agentic work, or when your evals on Opus 5 at higher effort still fall short.

That is a vendor telling you not to buy the more expensive product by default, which is worth weighting heavily. It also sets the decision procedure: Opus 5 is the baseline, and Fable 5.1 has to earn its place by beating it on your evaluation set, not on a launch narrative. The same relationship shows up in the fallback configuration &mdash; the permitted fallback targets for a refused Fable 5.1 request are Claude Opus 4.8 and Claude Opus 5.

The case for Fable is narrow but real. Its stated strengths are long-running agentic coding over sessions that run for hours, multi-file features, large refactors and migrations, multistep research that follows up on what it finds, vision on dense charts and filings, reasoning across the full 1M-token context window, and computer use.

Note the shape: these are all tasks measured over hours rather than single responses, and single-shot benchmarks capture them badly. If your workload is a long autonomous loop where a failure halfway through wastes everything before it, a higher completion rate can be worth double the token price. If your workload is a few hundred bounded requests a day, it almost certainly is not.

Both models also support changing effort mid-conversation, so the more interesting question is often not which model but which effort level. Anthropic's guidance points at Opus 5 at higher effort as the thing to beat before paying for Fable at all.

There is one place Fable 5.1 is genuinely cheaper, and it is counter-intuitive. Cache reads on Fable 5.1 cost 0.025 times base input, which is $0.25 per million. On other Claude models, including Opus 5, cache reads cost 0.1 times base input &mdash; which on Opus 5's $5 base is $0.50 per million. Fable 5.1 reads a cached prefix at half the price of Opus 5.

It does not flip the overall answer, because output tokens still cost double. Take 100 million input tokens with 90% served from cache, and 10 million output tokens. Fable 5.1 comes to roughly $622: $100 of fresh input, $22 of cached reads, $500 of output. Opus 5 comes to roughly $345: $50, $45 and $250. Opus stays cheaper by a wide margin, because output dominates.

Push the cache ratio to 99% and cut output to a tenth and the gap narrows to about $85 against $80 &mdash; close, but still not an inversion. The cache advantage is real and worth knowing; it is not a reason to pick Fable.

Before the model choice, there is a cheaper decision that gets skipped. Both Opus 5 and Fable 5.1 take an effort parameter controlling how deeply the model thinks, and on Fable 5.1 it defaults to high. Anthropic's migration guidance explicitly says to re-tune it from the default, which is an unusual thing for a vendor to say about its own default unless the default is deeper than most work requires.

Because thinking tokens bill as output tokens, effort is a direct multiplier on the most expensive line of your bill. Raising Opus 5's effort costs you a fraction of what moving to Fable costs, and the documentation frames the model switch as the thing you do only after Opus 5 at higher effort has already failed. Both models also allow changing effort mid-conversation without invalidating the prompt cache, so the sensible pattern is a low baseline with deliberate escalation on the hard steps, rather than one level held all session.

There is a perception trap worth naming. When you have chosen the premium option, its successes confirm the choice and its failures read as a hard problem rather than a weak model. Benchmarks exist precisely because that intuition is unreliable, and here the benchmarks point the other way while the price points at Fable.

Launch coverage compounds it. A new release generates a burst of confident comparisons written before anyone has run it on real work, and for Fable 5.1 specifically there were no published scores to write them from. Anything ranking Fable 5.1 against Opus 5 on capability in the launch window is an opinion wearing a table's clothes. Your own evaluation set, run on your own tasks, is the only thing that settles it &mdash; which is exactly what Anthropic's documentation tells you to do.

Take an agent consuming 100 million input and 10 million output tokens a month, with no caching. On Fable 5.1 that is $1,000 of input and $500 of output: $1,500. On Opus 5 it is $500 and $250: $750. The saving from picking the cheaper model is $750 a month, and on the published benchmarks you are also picking the higher-scoring one.

Now add caching at 90%, which is realistic for an agent that re-reads a stable prefix. Fable 5.1 falls to roughly $622 and Opus 5 to roughly $345. The ratio narrows slightly, because Fable's cache reads are half the price of Opus 5's, but the absolute gap stays near $280 a month. Output tokens are what decide this, and Fable charges double for them in every scenario.

• ▸Start on Opus 5. It is Anthropic's recommendation, it is half the price, and it leads every published benchmark.
• ▸If Opus 5 falls short, raise its effort level before changing model. That is the cheaper lever and it is the one the docs point at.
• ▸Move to Fable 5.1 only for long-horizon agentic runs, and only after your own evals show it winning. Launch coverage is not evidence.
• ▸If your workload is cache-heavy and output-light, re-run the arithmetic rather than assuming &mdash; Fable 5.1's cache reads are half of Opus 5's.
• ▸If the workload is high-volume and routine, the real question is not which frontier model but whether it needs one at all.

Both of these are hosted, closed-weight and metered. For a large share of everyday development work &mdash; boilerplate, renaming, format conversion, first-draft tests, summarising diffs, answering questions about a codebase &mdash; you are paying frontier prices for work an open-weight model handles acceptably on hardware you already own, at no cost per token.

A 32B-class coding model fits a 24 GB card at Q4_K_M with usable context; a 14B fits 16 GB comfortably; an 8B runs on 8 GB. None of them will out-reason Opus 5 on a hard problem, and none of them need to in order to absorb the volume that makes your bill. Our <a href="/models">Model Explorer</a> takes your GPU, RAM and backend and ranks what actually fits, and the <a href="/gpu">GPU pages</a> break it down card by card.

For the hosted side, our <a href="/blog/fable-5-1-vs-fable-5-what-changed">breakdown of what changed in Fable 5.1</a> covers the migration details, and <a href="/blog/fable-5-1-max-plan-usage-limits">why Fable empties a Max plan</a> covers the subscription mechanics.