Claude Fable 5.1 vs Opus 5: Is Anthropic’s Best Model Worth the Cost?

Claude Fable 5.1 creates an unusual buying decision.

On one independent benchmark, Fable 5.1 Max costs around 1.6 times as much per task as Opus 5 Max.

On another real-world coding-agent evaluation, Fable 5.1 costs less than Opus 5 while performing better.

Both results can be correct.

That is because Anthropic did not simply make Fable 5.1 “cheaper.” It left standard input and output pricing unchanged while cutting the cost of reading cached context by 75%. Whether that matters enormously or barely at all depends on what your workload actually spends tokens doing.

The short answer

For general knowledge work, writing, analysis and ordinary coding, Opus 5 remains the better-value default.

For long-running coding agents and other workloads that repeatedly reread large amounts of context, Fable 5.1 has become much more interesting. We would start those workloads by testing Fable 5.1 High, rather than automatically beginning with Opus.

Use Extra High or Max only when your own results show that the additional reasoning improves the finished task enough to justify the extra tokens.

The important number is no longer just price per million tokens.

It is cost per completed task.

---

What changed in Claude Fable 5.1?

Anthropic released Claude Fable 5.1 on September 1 as the successor to Fable 5 for demanding reasoning, long-running agentic coding, research and complex knowledge work.

The headline specifications remain substantial:

Fable therefore still carries twice Opus 5's ordinary input and output price.

The major pricing change is prompt caching.

Fable 5 previously charged $1 per million cached input tokens. Fable 5.1 charges $0.25.

Cache-write pricing stays unchanged at $12.50/M for a five-minute cache and $20/M for a one-hour cache.

Anthropic estimates that the change reduces a typical Fable workload by around 25% and highly agentic workloads by as much as approximately 45%.

That is not a 25–45% discount on the model.

It is an estimate of what happens when cached reads make up enough of the bill.

---

Fable 5.1 appears to be genuinely stronger

The capability improvement is not based only on Anthropic's launch charts. For generational context, see our earlier Claude Fable 5 and Mythos 5 analysis.

Artificial Analysis currently scores Fable 5.1 at 66 on its Intelligence Index at Max effort, the highest result it has measured.

Fable 5.1 also improves over Fable 5 across several reasoning, coding and agentic evaluations. Artificial Analysis measured 59.1% on Humanity's Last Exam, plus its strongest results to date on several coding and agent tasks.

There is an important qualification.

On Artificial Analysis' professional knowledge-work evaluations, Fable 5.1 and Opus 5 are much closer than the headline Intelligence Index suggests. Fable 5.1's lead on GDPval-AA v2 sits inside the confidence interval, while AA-Briefcase is effectively tied.

Fable performed better on analytical quality and rubric correctness.

Opus performed better on presentation quality.

So this is not:

Fable scores 66, therefore use Fable for everything.

It is evidence that Fable 5.1 raises Anthropic's capability ceiling.

The value question is separate.

---

The cost comparison gets weird

Artificial Analysis measured the following cost per Intelligence Index task:

On this workload, Fable 5.1 Max is approximately 20% more expensive than Fable 5 Max and around 1.6× the cost of Opus 5 Max.

That sounds like the cheaper-cache story has failed.

It hasn't.

Artificial Analysis estimates that the cache reduction actually saved around $1.40 per task. Without it, the same Fable 5.1 Max workload would have cost approximately $5.16.

The problem was elsewhere.

Fable 5.1 Max generated around 1.7 times as many output tokens as Fable 5.

And output still costs $50 per million tokens.

So Fable became much cheaper at rereading context while simultaneously becoming capable of spending enough additional output tokens to overwhelm some of that saving.

That's the first half of the story.

---

Then Cognition gets the opposite result

Cognition tested Fable 5.1 on FrontierCode 1.1 Extended, a workload designed around long-running coding agents.

Its results look completely different:

On that workload, Fable 5.1 is not more expensive than Opus.

It is cheaper.

And compared with Fable 5, Cognition measured a 54% reduction in cost per task.

The reason becomes obvious when Cognition breaks the task down.

Its representative Fable 5.1 coding task contains roughly:

• $1.07 of output
• $0.84 of uncached input
• $3.08 of cached input at the old rate

At Fable 5's cache-read pricing, that workload would cost roughly $4.99.

Cut the cache-read component from $3.08 to approximately $0.77, and the total drops to $2.68.

Nothing mysterious happened.

The coding agent spends an enormous amount of its workload rereading context it has already seen.

That is exactly the work Anthropic just made four times cheaper.

---

How can Fable be more expensive than Opus in one test and cheaper in another?

Because the workloads are different.

Artificial Analysis uses a broad basket containing reasoning, coding and knowledge-work evaluations. Fable 5.1's additional output-token use can matter heavily there.

Cognition is measuring an agent continually operating on a codebase, where cached input dominates the workload.

Artificial Analysis itself notes that Fable 5.1's cache savings were concentrated in its agentic evaluations, where most input tokens were cache reads.

This is the important pricing lesson:

Standard token price tells you what individual token types cost. It does not tell you what your completed job costs.

For a fresh one-shot task:

Fable's $10/$50 pricing can be expensive.

For an agent repeatedly reading the same large repository:

Fable's $0.25 cache reads can dominate the calculation.

That is why a single “Fable 5.1 is 25% cheaper” number is not very useful.

---

CursorBench strengthens the coding case

CursorBench evaluates agents on ambiguous, multi-file tasks taken from real Cursor engineering sessions.

Fable 5.1 currently holds the top two positions.

Cursor calculates cost using each model's published input, cache-read, cache-write and output rates. It also warns that small score differences may not always be statistically meaningful.

But the broad result matters.

Fable 5.1 Extra High beats Opus 5 Max on both measured coding quality and cost.

Fable 5.1 High scores between Opus 5 High and Max while costing considerably less than Opus Max.

And Fable 5.1 Max costs roughly 44% less per task than Fable 5 Max while scoring higher.

This is enough to change our recommendation for agentic coding.

---

Opus 5 should not automatically be your coding default anymore

Anthropic's general model guidance still recommends starting with Opus 5 for most workloads and moving to Fable 5.1 when Opus at higher effort falls short.

For broad knowledge work, we agree, consistent with our earlier Claude Opus 5 analysis.

For long-running coding agents, the evidence now points somewhere else.

General knowledge work

Start with Opus 5.

It costs half as much per standard input and output token, performs very close to Fable 5.1 on several professional-work evaluations, and is likely to be enough for most analysis, writing, document work and ordinary development.

Long-running agentic coding

Start by testing Fable 5.1 High.

Both Cognition and Cursor show the cheaper cache rate materially changing the economics of persistent coding workloads.

Extremely difficult agent work

Test Fable 5.1 Extra High or Max only where the incremental quality matters.

The benchmark winner is not automatically the economic winner.

---

Max is still a dangerous default

Fable 5.1 supports five effort levels:

Low → Medium → High → Extra High → Max

Anthropic recommends starting at High.

It describes Medium as the balanced setting, Low as the efficiency option, Extra High as appropriate for long-running agentic work, and Max as the unconstrained setting for tasks where maximum capability matters more than token consumption.

Anthropic also says developers should rerun their effort-level tests even if they already calibrated Fable 5 because the names do not map to identical amounts of reasoning between generations.

That makes blindly carrying a Fable 5 configuration over to 5.1 a mistake.

For example, CursorBench gets:

Fable 5.1 High: 69.4% at $4.80/task

Fable 5.1 Max: 73.4% at $9.64/task

The cost roughly doubles for four additional percentage points on that benchmark.

Sometimes those four points will be worth far more than five dollars.

Sometimes they will be buying an extremely expensive refactor of a button label.

---

[Anthropic's own prompting guide](https://platform.claude.com/docs/en/build-with-claude/prompt-engineering/prompting-claude-fable-5-1) explains some of the token burn

There is a slightly amusing tension in the release.

Anthropic says Fable 5.1's writing generally improves on earlier Claude models, with fewer stock phrases and less unexplained jargon.

But its own prompting guide also warns about several behaviors that can increase output usage.

At Extra High and Max, the model can spend substantially longer reasoning before producing a long deliverable. For some requests it may effectively draft material during reasoning and then produce the final version afterward.

Anthropic also specifically warns that Fable 5.1 can be more likely than Fable 5 to rewrite an entire file when a targeted edit would have been sufficient.

Its recommended fix is to explicitly ask for surgical edits where a full rewrite would not change the result.

That behavior matters when output costs $50/M.

A model can be more capable and still become worse value if the harness repeatedly asks it to spend that capability unnecessarily.

---

Claude Max users are reporting a different cost problem

API economics are only half the story. This distinction also matters in Choosely's earlier analysis of usage-based AI subscription pricing.

Since Fable 5.1 launched, multiple Claude Max users have reported that their five-hour and weekly usage allowances appear to disappear unusually quickly while using the new model.

One Max-user thread describes roughly half of a weekly Fable allowance disappearing in a single day. Another user reported approximately 30% of a five-hour quota disappearing during a single substantial task. The release discussion hub contains similar reports.

These are user reports, not controlled tests.

Some commenters say they have not noticed the same behavior, and the reports do not prove Anthropic reduced subscriber quotas.

But the underlying concern is plausible given what we know about the model's token behavior.

Artificial Analysis measured much higher output consumption from Fable 5.1 Max, while Anthropic itself recommends reducing effort and avoiding unnecessary whole-file rewrites where the additional compute does not improve the task.

The crucial point is that subscription usage and API pricing are different systems.

---

The 75% cache cut does not mean a Claude Max user gets 75% more Fable

Anthropic's cache-read discount is an API price.

A developer using the Claude API can see exactly how much a cached token costs and calculate the savings.

Claude Pro and Max subscribers instead operate under plan-level usage allowances.

We found no public Anthropic documentation that translates those allowance percentages directly into the API's $0.25 cache-read rate.

So these two statements can coexist:

A Fable 5.1 coding agent can become cheaper in API dollars.

A Claude Max subscriber can still feel that Fable 5.1 exhausts their plan allowance faster.

That is not inherently contradictory.

It means you should not take Anthropic's “up to approximately 45% cheaper” agentic-workload estimate and assume your Max subscription will suddenly last 45% longer.

Different meter.

Different economics.

---

Which model should you actually use?

Choose Opus 5 for:

Writing and general knowledge work.

The performance gap is often much smaller than the token-price gap.

Normal analysis and research.

If Opus already produces the right answer, paying for additional Fable reasoning buys little.

Short or fresh conversations.

If there is little repeated context, Fable's cache advantage has less room to matter.

Routine development.

Small fixes, explanation, straightforward refactors and ordinary code generation do not automatically need Anthropic's most capable model.

High-volume tasks.

Twice the standard token price becomes painful quickly when Fable produces only marginally better outcomes.

Choose Fable 5.1 for:

Long-running coding agents.

This is the strongest evidence-backed use case so far.

Large repositories with heavy context reuse.

The more often the model rereads the same context, the more valuable the new cache rate becomes.

Hard debugging or architecture work.

Especially when cheaper models have already failed.

Long-horizon research and agents.

Where failure, repetition or losing task state is more expensive than the additional model cost.

Work where a better result avoids another entire attempt.

A more expensive successful task can still be cheaper than two cheaper failed ones.

---

Which Fable 5.1 effort level should you use?

Our current starting point would be:

This is not a permanent leaderboard.

It is a sensible starting framework from the evidence available at launch.

The right answer is still to benchmark your job, not ours.

---

Choosely verdict

Claude Fable 5.1 is Anthropic's strongest model. Whether it is better value than Opus 5 depends more on the shape of the workload than the sticker price suggests.

For general analysis, writing, research, document work and ordinary coding, Opus 5 remains our default. Artificial Analysis measured Opus 5 Max at $2.34 per task against $3.76 for Fable 5.1 Max, while the models were much closer on several professional-work evaluations than their headline Intelligence Index scores imply.

For persistent coding agents and other cache-heavy workloads, the recommendation can reverse. Cognition measured Fable 5.1 at $2.68 per completed coding task versus $3.51 for Opus 5, while CursorBench currently has Fable 5.1 leading its agentic coding evaluation across several useful effort levels.

That gives us a more useful decision hierarchy than simply ranking the models:

General work → Opus 5

Agentic coding and cache-heavy work → Fable 5.1 High

Exceptionally difficult long-horizon work → test Fable 5.1 Extra High or Max

And whichever model you choose, measure cost per successful task, not simply price per million tokens.

Anthropic made rereading context dramatically cheaper.

It did not make every Fable workload cheaper.

The strongest model is not automatically the right model. And the cheapest token is not automatically the cheapest task.

---

Frequently asked questions

Is Claude Fable 5.1 better than Opus 5?

Fable 5.1 currently scores higher on several independent intelligence and coding evaluations. However, Opus 5 remains competitive on professional knowledge work and costs half as much per standard input and output token. The better model depends on the workload.

Is Claude Fable 5.1 cheaper than Opus 5?

Not universally. Fable 5.1 costs twice as much for standard input and output tokens. But its cache reads cost $0.25/M. In cache-heavy agent workloads, Cognition measured Fable 5.1 costing less per completed task than Opus 5.

Is Fable 5.1 cheaper than Fable 5?

Standard input, output and cache-write prices are unchanged. Cache reads are 75% cheaper. Anthropic estimates this reduces typical workload costs by around 25% and highly agentic workloads by up to approximately 45%, but actual savings depend on workload composition.

Why can Fable 5.1 still cost more despite cheaper cache?

Because cache reads are only one part of the bill. Artificial Analysis measured Fable 5.1 Max generating roughly 1.7 times as many output tokens as Fable 5 Max, causing total cost per task to rise despite substantial cache savings.

Which Fable 5.1 effort level should I use?

Anthropic recommends starting at High, moving down to Medium or Low where quality holds, and stepping up to Extra High or Max only for highly capability-sensitive work.

Should I use Fable 5.1 for coding?

For long-running agentic coding and large repositories, yes — it is one of the strongest current use cases. Cursor's evaluation currently ranks Fable 5.1 Max first, while Cognition reports that its new cache economics make Fable 5.1 cheaper per completed coding task than Opus 5 on its FrontierCode workload.

Why is Fable 5.1 using my Claude limits so quickly?

Some Max subscribers are reporting faster-than-expected usage consumption, but community reports do not prove Anthropic reduced plan allowances. Fable 5.1 can use substantially more reasoning and output at high effort, and Anthropic recommends tuning effort and avoiding unnecessary full-file rewrites where they do not improve results.

Your AI stack shouldn't go stale. Save the AI tools you actually use in Choosely and keep track of pricing, capability and model changes as they happen. Build your AI stack free →