Getting the answer right is only part of finishing a report. Someone still has to decide how much time and money that answer deserves. Anthropic’s new Claude Fable 5.2 makes that choice more interesting: independent tests show substantial capability, alongside sharply different costs depending on the job and the amount of thinking allowed.

Based on 14 reviewed sources · See sources

A specialist for the difficult jobs

Anthropic released Fable 5.2 and Mythos 5.2 on 9 November. They share an underlying model, with different safeguards: Fable is the broadly available version; Mythos access remains restricted to vetted organisations in programmes including cybersecurity and life sciences. The company still recommends Opus 5.5 for everyday work, escalating to the new models for harder problems.

The original announcement presents improvements in coding, reasoning and vision. Those are company claims. The more useful question for a business reader is where independent testing finds an advantage, and what it takes to obtain it.

From the source · Claude · Anthropic ·

Anthropic introduces Fable 5.2 and Mythos 5.2. Independent evaluations and practical conditions follow below.

View the original post on X ↗

A lead, with a setting attached

Artificial Analysis places Fable 5.2 first on its Intelligence Index v5, at 65 with maximum effort. That configuration averages $26.54 per index task. Its evaluation uses Anthropic’s default fallbacks, which can route certain requests to another Claude model. This is a composite test of several capabilities, not a measurement of how much better your next campaign report will be.

Vals also ranks Fable 5.2 first, at 71.31% on its different index using high effort, with default fallbacks. Sonnet 5.5 follows closely at 70.85%. The two evaluators’ scores use different tests and scales; 71.31 is not directly comparable with 65.

Crucially, Vals’ effort comparison found that turning Fable up to maximum made its overall result worse. Cost rose from $17.94 to $49.62 per task, with unfinished tasks hitting time, token or output limits. That finding belongs to Vals’ test conditions. It challenges the assumption that the most demanding setting is automatically the best choice.

When more thinking finds more fixes

A separate bug-fixing experiment shows the other side of that trade-off. Paweł Huryn, who runs The Product Compass newsletter, tested two software projects with 105 planted defects. His Bug Hunt Bench uses the same instruction for each project and scores the changes to the code, rather than accepting a model’s own report that it fixed something.

Fable 5.2 found and fixed more of those defects at maximum effort than at low effort, with a much larger estimated bill. Its three maximum-effort runs scored 56, 53 and 65. The spread matters: a mean is not a promise that the next run will produce the same result.

From the source · Paweł Huryn · The Product Compass ·

Huryn reports his original repeated bug-fixing test. The run conditions and estimated costs are distinguished below.

View the original post on X ↗

The retained run notes describe Grok, another AI model, grading the fixes without being told which model produced them. The answer key remains private. Costs are estimates from token counts at published rates, not subscription invoices. Earlier rival-model rows used different judges and software versions, so they do not provide a clean like-for-like ranking. This is useful evidence about one repair task, not a measurement of a team’s productivity.

The detail

More fixes, a larger estimated bill

Fable 5.2 in the same two-project, 105-defect task. Each row averages repeated runs.

More fixes, a larger estimated bill. Fable 5.2 in the same two-project, 105-defect task. Each row averages repeated runs.
Thinking settingMean defects fixed / 105RunsEstimated USD per run
Low412$24.32
Maximum583$139.65
Same model, Claude Code version and blind AI judge; maximum-effort runs sometimes compacted their conversation history. List-rate cost estimates include both projects. Private answer keys prevent a full independent reproduction; we did not rerun this test.Source: Bug Hunt Bench · retained run notes, 10 November · 2026-11-10

What one complete analytics run cost

Plotly’s new evaluation offers a business-shaped example. In Plotly Studio Embedded through OpenRouter, Fable 5.2 answered all 43 questions correctly, improving on 5.1. GPT-6.1 Sol also scored perfectly at a much lower cost. Plotly sells analytics software; this is its own test, separate from Anthropic’s launch demonstrations.

The run used more tokens than 5.1, and most of its bill came from writing the prompt cache: storing context for reuse. That helps explain how a lower unit price can still produce a higher total.

The detail

One exam, different bills

Plotly’s private, synthetic wind-energy dataset; 43 questions per model.

One exam, different bills. Plotly’s private, synthetic wind-energy dataset; 43 questions per model.
ModelCorrectWhole run, USDMedian per question
Claude Fable 5.243 / 43$35.7440 seconds
Claude Fable 5.139 / 43$31.4927 seconds
Claude Opus 5.541 / 43$13.3127 seconds
GPT-6.1 Sol43 / 43$2.6118 seconds
One full run per model in standard mode, with the same harness and grading. Prices are those when each model ran. A private synthetic exam does not establish a universal winner or a forecast for your bill.Source: Plotly · 10 November evaluation · 2026-11-10

Cheaper tokens, a separate subscription question

Anthropic’s model documentation lists standard API rates of $8 per million input tokens and $40 per million output tokens, 20% below 5.1. Tokens are the chunks of text a model processes or produces; reasoning and repeated context can add up. Optional fast mode doubles those unit rates. The API defaults to high effort, while Claude’s apps and Claude Code default to medium.

Fable is available on paid Claude plans, but access is not unlimited included usage. Pro and standard Team seats normally use separate pay-as-you-go credits. A temporary launch offer includes Fable within up to half their weekly allowance through 16 November at 11:59pm Pacific time — 17 November at 08:59 in Brussels. Other eligible plans have separate reset offers. After the promotion, the usual plan rules return.

The same distinction appeared in our Haiku 5.5 reporting: cheaper text processing and the time spent thinking are separate parts of the decision.

Before the client files go in

For European teams handling confidential client work, data terms are another practical gate. Anthropic’s Covered Models policy normally retains prompts and responses for at least 30 days for safety monitoring, with longer retention possible for investigations or legal requirements. Eligible, expressly authorised customers can have different arrangements; Enterprise Frontier Safeguards is in private beta. An existing zero-retention expectation should be checked against the actual account terms before using Fable.

Buy the result you need

The strongest conclusion is narrower than “replace everything”. Fable 5.2 has earned attention for demanding work. The independent evidence also makes the setting, the workflow and the definition of finished work part of the buying decision.

For a recurring analysis, compare the same brief and acceptance criteria: correct findings, useful explanation, time to a reviewed answer and the complete bill. A model that reasons for longer may rescue a difficult problem. It may also spend the budget without finishing. Both possibilities now have evidence behind them.

Sources

Written with AI assistance from the sources above. Analysis and practical implications are our interpretation. Our editorial approach.