🔍 Read the full analysis: Claude Fable 5.2: See How Effort Level Affects The Cost Of A Task on ThorstenMeyerAI.com
Prime made for students and young adults
- Fast, free delivery for dorm and study essentials
- Prime Video and Amazon Music included
- Member-only deals
TL;DR
Anthropic released Claude Fable 5.2 on November 9, 2026, with five selectable effort levels. In Artificial Analysis testing, scores ranged from 54 at Low effort to 65 at Max, while average cost per test task rose from $3.10 to $26.54. The results are a benchmark comparison, not a guarantee of the cost or performance businesses will see on their own work.
Anthropic released Claude Fable 5.2 on November 9, and benchmark results published by Artificial Analysis show a wide cost range depending on its effort setting: average cost per test task ran from $3.10 at Low to $26.54 at Max. For context, see the earlier Claude Fable 5.1 cost analysis. The top setting scored 65 on the Intelligence Index, while the least expensive scored 54, giving teams a concrete trade-off to weigh between performance, cost and wait time.
Artificial Analysis tested five settings—Low, Medium, High, Xhigh and Max—on its Intelligence Index, which covers nine evaluations in areas including agentic coding, business documents, science and knowledge. At Medium, the default in Claude apps and Claude Code, Fable 5.2 scored 56 for $3.80 per task. High scored 62 for $9.08, Xhigh also scored 62 for $14.83, and Max reached 65 for $26.54. These are the evaluator’s average API costs for one task in its test, not a universal price per user request.
The reported average task time also rose with effort: 6.5 minutes at Low, 8.3 at Medium, 17.9 at High, 26.4 at Xhigh and 46.4 at Max. Artificial Analysis reported that the model’s list price is the same across settings—$8 per million input tokens and $40 per million output tokens. The difference in test bills came from token use: the model generated 34.9 million output tokens across the index at Low and 275 million at Max.
The benchmark does not show that every task benefits from the most expensive setting. Artificial Analysis reported that High scored 49.1% on its FrontierCode 2.0 test, slightly above Max at 48.8%. Its rounded Intelligence Index score was also the same at High and Xhigh, 62, despite the higher Xhigh cost. Results depend on the task and evaluation; neither finding establishes that one setting is best for all users.
Claude Fable 5.2 Effort, cost and performance
Artificial Analysis found that higher effort generally raised benchmark scores, task costs and completion times. These results compare settings on its tests; they do not predict every team’s results or bill.
Effort brings trade-offs
The five settings give teams choices about quality, spend and waiting time.
Lowest test cost
Scored 54 and averaged $3.10 per task in this benchmark.
Default setting
The Claude apps and Claude Code default scored 56 at $3.80 per task.
Highest index score
Scored 65, with an average task cost of $26.54 and time of 46.4 minutes.
More effort is not always better value
Some results show why teams should match effort to the task.
- Intelligence Index score: 62
- Average cost: $9.08 per task
- FrontierCode 2.0 score: 49.1%
- Intelligence Index score: 65
- Average cost: $26.54 per task
- FrontierCode 2.0 score: 48.8%
Test it on real work
Benchmark averages are not a forecast of an organization’s performance or API bill.
Include the kinds of work your team expects Fable to handle.
Record quality, token use, cost and completion time.
Use a higher setting where its added value justifies the cost and delay.
The Cost of More Model Effort
The figures give organizations a way to think about effort as a budget and latency choice, rather than a setting to turn up automatically. In this benchmark, moving from Medium to High raised the score by six points and the reported cost by $5.28 per task. Moving from High to Max added three points, but cost nearly three times as much and took substantially longer on average.
That matters for teams running many requests or workflows that must return quickly. A higher score on a broad benchmark may be valuable for a difficult research or coding task, but the additional spend and delay may not be justified for routine work. The test results offer comparative evidence, not a forecast of a company’s savings or productivity: actual bills depend on the prompts, token use, workload and chosen model settings.
Artificial Analysis also compared Fable 5.2 at High with Anthropic’s Claude Opus 5.5 at Max. It reported scores of 62 versus 59 and average task costs of $9.08 versus $10.04, respectively. This is one evaluator’s result on its index; it does not establish that Fable will outperform Opus across other tasks or deployment conditions.
Products Worth Considering
Bose QuietComfort True Wireless Earbuds with Active Noise Cancellation, Bluetooth 5.3, USB-C Charging, IPX4, Black
LG UltraGear 27-inch QHD OLED Gaming Monitor, 240Hz
Apple AirPods Pro (2nd Generation) Wireless Earbuds with USB‑C Charging, Up to 2X More Active Noise Cancelling, Transparency Mode, Adaptive
As an affiliate, we earn on qualifying purchases.
Fable, Mythos and the Release
Anthropic’s November 9 release replaces Fable 5.1 and Mythos 5.1. The source account describes Fable 5.2 as available to users with safeguards for high-risk cybersecurity and most research biology tasks. It says Mythos 5.2 is offered to organizations accepted into Anthropic’s Cyber or Life Sciences Verification Programs, without those same safeguards. Anthropic says the two names refer to the same underlying model with different access and safety controls.
Anthropic said the new models produce answers up to 30% faster than the 5.1 versions and listed prices of $8 per million input tokens and $40 per million output tokens. It also described an optional fast mode, priced at $16 and $80 per million input and output tokens, that it says can generate tokens up to 2.5 times faster. Those speed claims come from Anthropic, not the Artificial Analysis effort-level table.
The source material reports that Anthropic recommends starting with Claude Opus 5.5 for most coding and knowledge work, and moving to Fable when a task calls for it. Anthropic also described striking results from prerelease tests, including a large multi-agent attempt to formalize Fermat’s Last Theorem. Those are vendor-reported demonstrations, not independent evidence of what typical users will achieve. The cost-and-score comparisons in this report come from Artificial Analysis, whose model pages were accessed November 10.
“Start with Claude Opus 5.5 for most coding and knowledge work, and move up to Fable when the task needs it.”
— Anthropic
What the Benchmark Cannot Show
The published figures do not establish how Fable 5.2 will perform on any particular organization’s workload. Artificial Analysis’s cost per task reflects its own Intelligence Index tasks and API-cost calculations; readers should not treat it as a quoted price for every prompt or a prediction of an individual team’s bill.
The source material does not provide the full task-level results, uncertainty ranges or enough detail to determine how consistently each effort setting performs across repeated runs. Index scores are rounded, so High and Xhigh sharing a score of 62 does not prove their underlying performance is identical. Anthropic’s claims about faster generation, reduced safety blocks and prerelease demonstrations also have not been independently verified in the information provided.
Testing Effort on Real Workloads
Organizations considering Fable can compare settings on representative tasks and record quality, token use, cost and completion time before broad deployment. That would help identify whether a higher effort level adds enough value for a particular workflow, rather than assuming the top benchmark score is worth its additional expense.
Further independent testing could clarify how the model performs across repeated runs, practical workloads and safety-sensitive tasks. Until those results are available, the November 10 figures describe a specific benchmark snapshot, while access to Mythos remains limited under Anthropic’s stated verification programs.
Key Questions
What did Artificial Analysis find about Fable 5.2’s effort settings?
It reported Intelligence Index scores from 54 at Low to 65 at Max. Its average cost per test task ranged from $3.10 to $26.54, with higher effort generally taking longer.
Is Max effort always the best setting?
No. Max had the highest score on this index, but it also had the highest reported cost and average time. On Artificial Analysis’s FrontierCode 2.0 test, High scored 49.1%, slightly above Max at 48.8%.
What does the reported cost per task mean?
It is Artificial Analysis’s average API cost for a task in its Intelligence Index evaluation. It is not a fixed fee for every user request; real costs depend on workload and token use.
How are Fable 5.2 and Mythos 5.2 different?
The source material describes them as the same underlying model with different access and safeguards. Fable is available with safety controls, while Mythos is restricted to organizations accepted into Anthropic’s relevant verification programs.
Source: ThorstenMeyerAI.com
