Anthropic unveiled Claude Fable 5.2 on November 9, claiming superiority over Claude Opus 5.5 across nearly all measured capabilities. Yet the company simultaneously advised developers to begin with Opus for most coding and knowledge work, describing it as its "strongest daily driver." This apparent contradiction prompted a closer examination of whether the more expensive model justifies its cost.
Anthropic's own benchmarks show Fable 5.2 achieving 49.1% on FrontierCode 2.0 at Xhigh effort versus 45.5% for Opus 5.5 at Max. On FrontierSWE v2, Fable reached 69.9% compared to Opus's 62.3%. However, Fable's standard API pricing runs at twice the rate: $8 per million uncached input tokens and $40 per million output tokens, against Opus's $4 and $20 respectively.
To evaluate whether the price premium delivers tangible benefits, three practical tasks were executed on both models using Anthropic's API at default settings. Each model ran once per test without modification to standard configurations. Fable 5.2 operates at high effort by default, while Opus 5.5 defaults to medium effort, reflecting what developers receive without additional tuning. A grading script evaluated each response, with API costs calculated from reported token consumption across uncached input, cache reads, cache writes, and output.
Test Scenarios
- Cron scheduler: Models received a specification for scheduled job execution code, followed by a modification request. A hidden suite of 50 tests evaluated the final implementation.
- Invoicing repository fix: Models operated as agents within a small invoicing library containing interacting bugs in rounding, discounts, tax, and shipping calculations, plus one intentionally incorrect test.
- Incident investigation: Models used terminal access to identify the root cause of a payments outage using logs, metrics, and deployment records, including misleading deploys and servers with incorrect system clocks.
Both models achieved perfect scores on initial test versions, prompting iterative increases in difficulty with two additional rounds per scenario.
Cron Scheduler Results
This task assessed whether a model could translate a detailed specification into functional code and incorporate subsequent changes without introducing errors. Test complexity evolved across three versions:
- Version 1 covered standard cron schedules such as 15-minute intervals or weekday mornings at 9 a.m.
- Version 2 introduced complex rules including "the weekday nearest the 15th" and "the last Friday of the month," with a change request that reversed field interaction logic.
- Version 3 required manual time zone and daylight saving time handling without Python's built-in time zone library.
Both models scored 50 out of 50 across all three versions. The distinction lay in computational overhead: Fable 5.2 generated two to three times more reasoning tokens than Opus 5.5 on every iteration, yet this additional reasoning produced no score improvement.
The extra reasoning did not raise Fable's score on any of the three versions.
Across the three scheduler iterations, Opus completed the work in 358 seconds total at an estimated cost of $0.99, while Fable required 661 seconds and cost $3.29.
- Version 1: Fable 5.2 used 14,471 input and 18,732 output tokens (11,122 reasoning) in 170 seconds for $0.87; Opus 5.5 used 8,055 input and 9,431 output tokens (3,846 reasoning) in 80 seconds for $0.22.
- Version 2: Fable 5.2 used 20,215 input and 23,990 output tokens (14,627 reasoning) in 226 seconds for $1.12; Opus 5.5 used 14,090 input and 15,991 output tokens (6,695 reasoning) in 131 seconds for $0.38.
- Version 3: Fable 5.2 used 21,121 input and 28,254 output tokens (17,452 reasoning) in 265 seconds for $1.30; Opus 5.5 used 12,651 input and 17,205 output tokens (6,533 reasoning) in 147 seconds for $0.39.
Invoicing Repository Fix Results
This debugging exercise mirrors work agents perform for developers on a regular basis. Test iterations introduced increasing complexity:
- Version 1 contained bugs in rounding, discounts, tax, and shipping, plus one incorrect test case.
- Version 2 added buy-two-get-one promotions, weight-based shipping, and refunds, each with associated bugs.
- Version 3 introduced currency conversion and invoice numbering, again with bugs embedded.
This scenario produced the most pronounced divergence in cost and execution time. Both models achieved perfect 50 out of 50 scores across all versions, fixing all bugs the grader assessed and correctly identifying the erroneous test without modifying it. Opus 5.5 consumed fewer reasoning tokens on every version and completed each iteration in under one minute, while Fable 5.2 required between one minute 46 seconds and just under four minutes, costing eight times more overall across the three runs.
Despite Fable's additional reasoning capacity, no score improvement materialized.
- Version 1: Fable 5.2 used 17,585 input and 11,511 output tokens (5,909 reasoning) in 106 seconds for $0.58; Opus 5.5 used 10,585 input and 2,975 output tokens (519 reasoning) in 28 seconds for $0.09.
- Version 2: Fable 5.2 used 88,697 input and 25,980 output tokens (11,820 reasoning) in 223 seconds for $1.39; Opus 5.5 used 13,938 input and 4,457 output tokens (777 reasoning) in 39 seconds for $0.13.
- Version 3: Fable 5.2 used 29,660 input and 24,020 output tokens (14,022 reasoning) in 214 seconds for $1.18; Opus 5.5 used 24,208 input and 6,013 output tokens (1,178 reasoning) in 51 seconds for $0.17.
Incident Investigation Results
This task emphasized methodical analysis over rapid coding, suggesting potential advantages for Fable 5.2. Test difficulty increased as follows:
- Version 1 included misleading deploys and one server with a seven-minute clock drift.
- Version 2 added a clock correction mid-incident, request retries, a compressed log file, and duplicate log entries.
- Version 3 introduced a server logging in Chicago time and an unrelated error spike within the incident timeframe.
Both models achieved perfect scores (50 out of 50) across all three versions. Fable 5.2 required fewer conversational turns on every version—9 to 11 compared to Opus 5.5's 12 to 14. Across all three iterations, Fable's estimated cost reached $2.54 versus Opus's $0.91.
- Version 1: Fable 5.2 used 155,925 input and 14,309 output tokens (8,638 reasoning) in 143 seconds for $0.94; Opus 5.5 used 111,618 input and 7,405 output tokens (2,801 reasoning) in 73 seconds for $0.25.
- Version 2: Fable 5.2 used 168,913 input and 9,416 output tokens (3,551 reasoning) in 105 seconds for $0.68; Opus 5.5 used 188,846 input and 7,218 output tokens (2,518 reasoning) in 82 seconds for $0.32.
- Version 3: Fable 5.2 used 205,601 input and 11,887 output tokens (4,474 reasoning) in 127 seconds for $0.92; Opus 5.5 used 253,958 input and 7,598 output tokens (2,041 reasoning) in 70 seconds for $0.34.
Overall Assessment
Performance scores remained identical across all nine test iterations per model. Fable 5.2 incurred nearly four times the cost and required more than double the execution time.
These findings support beginning with Opus 5.5 for tasks of this nature. It matched Fable 5.2's performance and executed faster on every test version at approximately one-quarter of the overall estimated cost.
This comparison does not diminish Fable 5.2's capabilities. The test scenarios were challenging but limited in scope. Anthropic positions Fable for more exploratory technical work. Since each model ran at its API default configuration, the results leave unresolved how much of the performance gap stems from their different effort level settings.
A practical approach would involve starting with Opus 5.5 and transitioning to Fable 5.2 only when Opus fails to deliver satisfactory outcomes.