Anthropic's Claude Fable 5.2 cuts refusals sharply but delivers modest cost savings in real-world tests
Head-to-head coding benchmarks reveal that Fable 5.2 eliminates safeguard blocks that plagued its predecessor, though the promised 20% price reduction yields only marginal savings in practice.

Anthropic unveiled Claude Fable 5.2 with claims of reduced content filtering and lower API pricing, yet independent testing shows the improvements manifest unevenly across different development tasks. The company reported that Fable 5.2 triggered roughly 55% fewer safeguard blocks than Fable 5.1 in early testing, with cybersecurity scenarios showing a 60% reduction. The model also generates output up to 30% faster and carries standard API rates of $8 per million uncached input tokens and $40 per million output tokens—both 20% lower than Fable 5.1's $10 and $50 pricing.
To evaluate these claims, I ran three developer-focused tasks comparing the two models: a 50-turn webhook relay chat, a signature verification code challenge, and a login library debugging exercise. Each test tracked refusals, accuracy, token consumption, estimated API costs, and execution time. The comparison used a single run per model per task, with API calls configured at Anthropic's default High effort setting and no turn cap or fallback model. A custom grading script evaluated outputs against predetermined answer keys, and cost estimates account for Anthropic's separate pricing tiers for uncached input, cache reads, cache writes, and output.
The tests
Webhook chat
A 50-turn scripted conversation simulated a developer building a small webhook relay, incorporating 25 project rules—eight of which changed partway through and three that remained only implicit. The model was expected to write configuration and code, then answer an eight-question recall quiz. In an October run on Fable 5.1, the API refused 34 of 50 turns, with the first refusal occurring on turn 5 when asked to read the relay's target URL from an environment variable. That earlier run produced no code or quiz answers and scored zero on both components.
Fable 5.2 completed the identical script without any refusals. The model's code incorporated the most recent version of all eight rules that had been modified during the conversation, and on the quiz it answered "unknown" to the three questions lacking a definitive answer rather than filling gaps with speculation.
- Fable 5.2: 50 turns, 298,709 input tokens, 41,828 output tokens (28,550 reasoning), $4.06 estimated cost, 427 seconds
- Fable 5.1 (October): 34 of 50 turns refused, no code or quiz answers generated
Signature checker
This test provided a written specification for code that verifies signed webhooks, rejects replay attacks, and rate-limits senders, followed by a change request. A hidden test suite of 50 cases graded the final code output. Fable 5.1 refused both turns in under a second each, with the API labeling both refusals as "cyber." No output was returned on either attempt. Anthropic has noted that benign security work can trigger its cyber classifier.
Fable 5.2 accepted the same specification, processed the change request, and passed all 50 hidden tests, including edge cases involving timestamp handling, rotated secrets, and replay scenarios.
- Fable 5.2: 2 turns, 21,299 input tokens, 26,167 output tokens (18,407 reasoning), $1.22 estimated cost, 253 seconds
- Fable 5.1: 2 turns, both refused, 3,206 reported input tokens, 0 output tokens, $0.00 (Anthropic bills cyber refusals before output generation), 2 seconds
Login repo fix
Both models completed this task, allowing a direct performance comparison once refusals were eliminated. The test placed each model in an agent role within a small login library containing interacting bugs across password hashing, lockouts, reset tokens, and sessions. The library also included one incorrect test that the model should identify but not modify. Since both models achieved perfect scores initially, the test was made progressively harder and rerun twice.
- Version 1: Four modules with bugs in password hashing, lockouts, reset tokens, and sessions, plus one wrong test
- Version 2: Added a one-time-code module with its own bugs; tightened existing rules to require usernames to match across Unicode forms and sessions to handle a clock running backward
- Version 3: Added an audit log requiring secret redaction and a rate limiter, neither with pre-existing tests
Both models achieved 50/50 scores on all three versions, fixing the bugs covered by the grader and reporting the wrong test without editing it. Fable 5.1 never refused on the login repo fix, despite the task involving passwords, lockouts, and reset tokens.
Across the three versions, dividing total output tokens (including reasoning) by total elapsed time yields approximately 107 tokens per second for Fable 5.2 and 82 for Fable 5.1. Fable 5.2 completed each version in four turns, while Fable 5.1 required 7, 21, and 25 turns respectively. Fable 5.2 also generated more reasoning tokens on each version. On the first two versions, Fable 5.2 finished faster and at lower estimated cost; on the third, it took longer and cost more.
Version 1
- Fable 5.2: 4 turns, 16,149 input tokens, 4,502 output tokens (1,178 reasoning), $0.27 estimated cost, 43 seconds
- Fable 5.1: 7 turns, 33,935 input tokens, 4,761 output tokens (1,149 reasoning), $0.36 estimated cost, 52 seconds
Version 2
- Fable 5.2: 4 turns, 24,081 input tokens, 10,024 output tokens (4,887 reasoning), $0.56 estimated cost, 95 seconds
- Fable 5.1: 21 turns, 139,189 input tokens, 9,475 output tokens (4,018 reasoning), $0.68 estimated cost, 118 seconds
Version 3
- Fable 5.2: 4 turns, 34,269 input tokens, 17,964 output tokens (10,578 reasoning), $0.96 estimated cost, 166 seconds
- Fable 5.1: 25 turns, 208,826 input tokens, 10,638 output tokens (4,342 reasoning), $0.79 estimated cost, 132 seconds
Results
The combined token, cost, and time figures below encompass the signature checker and all three login repository versions. Fable 5.1 generated no output on the signature checker, while Fable 5.2 spent 253 seconds and an estimated $1.22 completing it. Across the three login versions that both models finished, elapsed time and estimated cost remained comparable. The combined totals reflect only the work that Fable 5.2 completed on the signature checker.
What do I think?
The most significant improvement demonstrated by Fable 5.2 in these tests was the elimination of refusals. Fable 5.1 blocked routine webhook and signature-checking tasks, whereas Fable 5.2 handled both without a single refusal. Anthropic maintains that the newer model continues to employ cyber safeguards.
In my tests, the biggest improvement with Fable 5.2 was fewer refusals.
Across the three login repository versions, Fable 5.2 produced roughly 30% more output tokens per second of elapsed time, including reasoning, while consuming approximately the same total duration. The combined runs required 304 seconds on Fable 5.2 and 302 seconds on Fable 5.1.
The displayed cost estimates totaled $1.79 for Fable 5.2 and $1.83 for Fable 5.1, representing a savings of about 2% across these runs. Anthropic's 20% price reduction applies to its standard per-token rates. Since each version ran once per model, testing both on your own workload would provide a clearer picture of the actual numbers for your use case.