Claude Fable 5.1 vs Fable 5: What Changed

‍

Claude Fable 5.1 is a refinement of Fable 5, and the difference shows up most in long, multi-step work rather than one-shot answers. The honest short answer is that Fable 5.1 is the better version of the two, at the same base price, but neither is the right model for every task. Anthropic spent weeks fueling Fable 5.1 speculation before the launch confirmed most of it; this guide covers what changed, what stayed the same, and how to decide, without assuming you write code.

Fable 5.1 and Fable 5 are the same model with different tuning

Fable 5.1 and Fable 5 come from the same underlying model, so this is a refinement, not a new generation. Anthropic released Claude Fable 5 on June 9, 2026, and Fable 5.1 on September 1, 2026, less than three months later. That short gap tells the story: Fable 5.1 keeps the identity of Fable 5 and sharpens the parts that decide whether a long task actually finishes.

The context window did not grow. The maximum output did not grow. The standard input and output prices did not change. Read a spec sheet alone and you might wonder why the decimal point was worth a launch.

The point of the update is follow-through. Fable 5 could take on an ambitious job; Fable 5.1 is built to stay oriented for more of it. That matters if you use Claude to run projects, not just answer questions.

What actually changed: cheaper cache reads and stronger long work

Two things changed that matter, and one headline number that looks bigger than it is. Cache reads got 75% cheaper, and the model got measurably better at long, agentic work. The "25% cheaper" line you may have seen needs a caveat, which we cover below.

The table below sorts what moved from what held steady, so you can see the shape of the update at a glance.

What changed and what stayed the same, Fable 5 to Fable 5.1 (as of September 2026)

Cache reads are what you pay when the model re-reads context it has already processed, such as a large set of instructions or a document it keeps referring back to. On Fable 5, that cost $1.00 per million tokens. On Fable 5.1, it is $0.25 per million tokens, a 75% cut. Anthropic estimates this lowers total cost by roughly 25% for a typical workload and up to about 45% for a highly agentic one, where re-reading the same context dominates the bill.

The rest of the price is unchanged. Input stays at $10 per million tokens, output stays at $50 per million tokens. Output is the expensive part of most bills, so if your work produces a lot of new text rather than re-reading stored context, the cache saving matters less.

The benchmark gains cluster in agentic and research work

Fable 5.1's biggest measured gains are in long, tool-using tasks, not in general reasoning. That pattern is consistent across Anthropic's own launch numbers: the harder and longer the task, the wider the gap.

The table below shows Anthropic's reported results, including Opus 5 for context. These are vendor figures, measured by Anthropic with its production safeguards on, so treat them as directional rather than independent proof.

Anthropic-reported benchmark results, Fable 5.1 versus Fable 5 and Opus 5 (as of September 2026)

The spread within the table is the real signal, not any single row. On Terminal-Bench-Science, which measures whether a model can carry a scientific research task through many terminal steps without losing the thread, Fable 5.1 more than doubles Fable 5's score. On CursorBench, a coding benchmark it already handled well, it moves only three points. Fable 5.1 is not twice as smart; it is much better where coordination and persistence over a long task matter, and only a little better on work Fable 5 already did competently.

Two of these deserve a plain-English translation, since they carry most of the "why upgrade" weight:

• AutomationBench measures business-workflow automation, the kind of multi-step office task where the model has to inspect information, decide what to do, and act rather than follow a fixed script. Nearly doubling here, from 17.1% to 31.4%, is the clearest sign Fable 5.1 is steadier at the messy, decision-heavy work that real tools demand.
• GDPval-AA v2 measures knowledge work, scored as an Elo rating rather than a percentage. Fable 5.1's 1,853 leads both Fable 5 and Opus 5, though Opus 5 sits close behind at 1,824.

Read the numbers with care. Anthropic notes a standard error of roughly 3.5 to 4.5 points on Terminal-Bench-Science, so narrow gaps should not be over-read, and every figure here was measured with production safeguards on, which Anthropic says likely lowers some scores where a safeguard interrupted a task.

Independent testing points the same direction. Artificial Analysis, which evaluates models separately from vendors, placed Fable 5.1 at the top of its Intelligence Index at 66 on max effort, the highest score it had measured, ahead of Opus 5 at 63 and Fable 5 at 62. Vendor benchmarks alone can flatter a model; independent corroboration on the same direction is what makes the pattern credible.

Pricing is the same on paper but the per-task story is subtler

Cheaper cache reads do not always mean a cheaper task, and this is the part most coverage skips. Fable 5.1 can cost more per finished task than Fable 5, even with the 75% cache cut, because it tends to use more output tokens to do the work.

1. The sticker price is identical except for cache reads

Base pricing did not move. Input stays at $10 per million tokens and output at $50 per million tokens, the same as Fable 5. The only price that changed is cache reads, cut 75% to $0.25 per million tokens. So on a line-item basis, Fable 5.1 looks strictly cheaper than Fable 5, never more expensive.

2. A finished task can still cost more, because 5.1 writes more

The per-task total can run higher despite the cheaper cache reads. Artificial Analysis measured this directly: across its Intelligence Index, Fable 5.1 at max effort cost $3.76 per task, against $3.14 for Fable 5, roughly 20% more. The reason is output. Fable 5.1 used about 1.7 times the output tokens to do the same work, and output is the $50 line. The cheaper cache reads did help, saving an estimated $1.40 per task; without that cut, the same run would have cost around $5.16. So the cache reduction is real, but on this workload heavier output more than absorbed it.

That is the whole paradox in one line: the cache cut pulls cost down, and higher output pushes it back up. Which effect wins depends on the shape of your work. A task that re-reads a large, stable context many times leans on cache reads, so 5.1 comes out cheaper. A task that generates a lot of new text leans on output, so 5.1 can come out more expensive. On Emergent, where you pay in credits rather than per token, the same mechanics sit underneath: a longer, more thorough run burns more credits.

3. Judge cost by the finished result, not the token price

Cost per completed task is the number worth watching, not price per token. A model that costs a little more per run but finishes in fewer attempts, with fewer dead ends, is cheaper overall than one that stalls and retries. That is why the honest comparison is per completed task, and it is the figure to keep an eye on as you build.

Cheaper cache reads lower the cost of re-reading context. They do not lower the cost of a model that writes more to get the job done. Watch the completed-task cost, not the headline rate.

What you will notice when a build runs on Fable 5.1

The differences you will actually feel are about behavior, not specs. Anthropic changed how Fable 5.1 works through a long task, and a few of those changes surface even when nothing about your prompt changes.

The ones that matter for building an app:

• Better follow-through on long jobs: Fable 5.1 holds a plan across many steps more reliably, so a complex build is less likely to drift or lose track of an earlier decision partway through.
• Denser, more efficient output: it tends to work with less back-and-forth chatter, though at higher effort it can also produce more, which is part of the per-task cost story above.
• Steadier recovery: when a step fails, it is more likely to diagnose and fix the root cause rather than paper over it, which matters most on multi-hour builds.

Under the hood, Anthropic also made developer-facing API changes in Fable 5.1, such as how forced tool use and internal reasoning are handled. If you build through a no-code platform, you do not touch any of that: the platform manages the model connection, so these changes affect how the build behaves, not anything you configure. If you ever export your app's code and wire it to Anthropic's API directly, that is the point to read Anthropic's migration notes.

Every figure has a provenance, and it changes how much weight it carries. Our Claude Fable 5.1 benchmarks guide labels each score as vendor-run or independently run.

Which one should you pick

Pick Fable 5.1 over Fable 5 whenever the choice is between those two, because it is the better-behaved version at the same base price. Fable 5.1 carries the newer tuning, the cheaper cache reads, and the newer knowledge cutoff. There is little reason to choose Fable 5 for new work unless you are matching an older project that already runs on it.

The more useful decision is when to reach for a Fable model at all. Fable is Anthropic's tier for demanding, long-horizon work, and it is not the economical default for everything. Anthropic's own guidance is to start with Claude Opus 5 for most workloads and move up to Fable only when a cheaper model keeps falling short on the hardest tasks.

Match the model to the job, not to the version number. The table below maps common project types to a sensible starting model.

Read the table as a starting point, not a rule. The honest test is whether a cheaper model finishes the job cleanly. If Opus 5 gets a complex CRM 90% of the way and keeps stumbling on the last stretch, that is exactly the kind of project to start on Fable 5.1 instead. If it sails through, you have saved money by not reaching for the premium tier out of habit. If you are weighing the Opus tier itself, our Opus 5 vs Opus 4.8 breakdown covers what changed there.

Both versions assume a Fable model is the right tier, and the 18-day suspension in June argued otherwise. Our best Claude Fable 5 alternatives guide covers 10 options and why a routing strategy beats a single default.

From choosing a model to building with it

The choice between Fable 5.1, Fable 5, and Opus 5 comes down to the job, not the version number: Fable 5.1 for your hardest long-running work, Opus 5 for everyday builds, and Fable 5 only when you are matching an older project. You do not have to lock that decision in ahead of time, though, or pick one model for everything.

On Emergent, all three are selectable, and you choose one per project when you start. Each runs through the Universal LLM Key with billing in credits, so you describe what you want in plain language and ship a working full-stack app from it, no API keys and no code. Match the model to what you are building, and start building on Emergent.