Claude Fable 5.1 vs Opus 5: Which One Earns Its Price

Fable 5.1 landed on 1 September 2026 and the obvious question is whether to move off Opus 5. The sticker price says no. The benchmarks say it depends entirely on what you run.

Here is how I would decide, with the numbers that matter rather than all nine.

The price gap is exactly double

Opus 5 is 5 dollars per million input tokens and 25 per million output. Fable 5.1 is 10 and 50. Anthropic prices in dollars, so those are dollars.

That is a clean 2x on both sides, and it has not changed with this release. What changed is underneath: cache reads on Fable 5.1 are now 0.25 dollars per million, down from 1.00. Anthropic puts the practical effect at roughly 25 percent off a typical workload and up to about 45 percent off a heavily agentic one.

So the honest comparison is not 2x. If your workload is cache-heavy, and long agent runs usually are because the system prompt and tool list repeat on every turn, the effective gap is meaningfully smaller than the price card implies. If your workload is a stream of unrelated one-off requests with no shared prefix, the cache cut does nothing for you and you are paying the full 2x.

Work out which one you are before reading any benchmark. It changes the answer more than the scores do.

Where the two models are effectively tied

On tasks that are one hard question rather than a long campaign, the difference is close to noise.

Humanity's Last Exam with tools: 65.0 percent for Fable 5.1 against 63.6 for Opus 5. That is 1.4 points. Without tools the gap is larger at 60.9 against 56.6, but still nothing like the price difference.

OSWorld 2.0 under strict scoring: 41.7 against 39.6. Two points.

GDPval-AA v2: 1853 against 1824. On a scale where the previous generation already scored 1723, that is under two percent.

For anything shaped like classification, extraction, summarisation, question answering, or a chat feature, paying double for two points is difficult to defend. Opus 5 is the value pick and it is not close. I made the same argument for the previous generation in Claude Opus 5 vs Fable 5: Which Tier Earns the Money, and 5.1 does not overturn it for this class of work.

Where the argument flips hard

Then there is Terminal-Bench-Science 0.1: 52.6 percent for Fable 5.1 against 29.0 for Opus 5. That is 23.6 points, and Fable 5.1 scores nearly double.

AutomationBench tells the same story with different numbers: 31.4 against 26.9, which is a relative improvement of about a sixth on a benchmark where both models are still failing most of the time.

Terminal-Bench 4.0: 55.8 against 52.3.

The pattern is not subtle once you sort by it. Every benchmark measuring a long autonomous run with many tool calls shows a real gap. Every benchmark measuring a single response shows almost none.

That is a coherent story rather than a marketing artefact. Long agentic work compounds: a model that makes slightly better decisions at each of forty steps does not end up slightly ahead, it ends up in a different place entirely. A single response has no compounding to do.

The decision, stated plainly

Stay on Opus 5 if your requests are short, independent, and mostly one round trip. You are buying two benchmark points for double the price and the cache discount cannot rescue that.

Move to Fable 5.1 if you run agents that work unattended across many tool calls, especially where a wrong turn early ruins the whole run. The gap there is large enough that the cheaper model is often the more expensive one once you count retries and human correction time.

Test both before committing either way, because there is a cheaper experiment most people skip. Lower effort on the newer model frequently beats higher effort on the older one. Before you build a routing layer that sends easy work to Opus 5 and hard work to Fable 5.1, try Fable 5.1 at low or medium effort on everything and measure. One model means one cache namespace, and a cascade forfeits cache reuse across its models, which can quietly eat the saving the cascade was built to capture.

Judge on cost per finished task, not cost per request. A cheaper request that needs a second attempt is not cheaper.

The one you cannot buy

Worth knowing before you compare tiers: there is a third model in this release that is not on the menu.

Mythos 5.1 is the same underlying model as Fable 5.1 with different safeguards, and it scored 60.9 percent on Terminal-Bench 4.0 against Fable 5.1's 55.8. That is a wider gap than the one between Fable 5.1 and Opus 5 on the same benchmark.

You almost certainly cannot get it. Access runs through the Cyber Verification Program and the Life Sciences Verification Program, and it is currently limited to a set of US organisations. So the top of the published table is not a purchasing option for a studio in Berlin, and any comparison that includes it is describing a model you cannot call.

Mention it only because the benchmark charts include it and it is easy to read those numbers as available. They are not.

What neither benchmark table tells you

Two practical differences matter more than a couple of points.

Fable 5.1 can decline. It returns HTTP 200 with `stop_reason: "refusal"` and a category in `stop_details`, and code that reads `content` without checking gets an empty string. Anthropic ships a server-side fallback parameter that routes those to another model automatically. Wire it in when you migrate, not after a user reports a blank screen.

Thinking is always on. You cannot disable it, and a request that tries returns a 400. Depth is controlled through `output_config.effort` instead, which takes low through max. If you are carrying forward a config that disabled thinking on an older model, it will not survive the move.

Anthropic also states its own blind spot: the behavioural audit has less visibility into very long-context work and multi-agent settings. Those are precisely the workloads the agentic scores recommend Fable 5.1 for, so the testing is thinnest exactly where the model is strongest. Bring your own evals.

A worked example, because percentages hide the money

Take an agent that runs 200 times a day. Each run carries a 15,000 token system prompt and tool list that never changes, then does 30 tool calls, and produces 3,000 output tokens.

The stable prefix is the whole point. On the first turn it is a cache write. On the following 29 it is a cache read, so per run you are reading roughly 435,000 cached tokens rather than paying full input price for them.

On Fable 5 that cache read cost 1.00 dollar per million, so those reads were about 0.44 dollars a run. On Fable 5.1 the same reads cost 0.25, or about 0.11. Across 200 runs a day that is a difference of roughly 66 dollars a day on the cache line alone, before any change in how many tokens the model needs to finish the job.

Now compare against Opus 5 honestly. Opus 5 is half price on input and output, so on the uncached portions it wins. But if Opus 5 needs a second attempt on even one run in ten because it took a wrong turn at step twelve, the retry costs a full run, and the arithmetic reverses.

That is why cost per finished task is the only number worth tracking. Cost per request flatters whichever model fails more often, because a failed run is cheap right up until you count the second one.

The figures above are illustrative arithmetic on public list prices, not a bill I am reading off a dashboard. Run the same shape against your own traffic before you decide.

Bottom Line

The two models are not competing for the same job, which is why "which is better" has no useful answer.

Opus 5 is the right default. It handles the overwhelming majority of real production traffic at half the price and within a couple of points on every benchmark that measures a single response.

Fable 5.1 earns its premium in one specific place: long, unattended, tool-heavy runs where compounding errors are expensive. There the gap is 23.6 points, not two, and the cache read cut takes a real bite out of the price difference on exactly the workloads that benefit.

The full nine-benchmark table, including the competitor columns, is in Claude Fable 5.1 Benchmarks. The rest of the model coverage lives in the Lab Overview.