Anthropic unveiled Claude Fable 5.2 with claims of substantially fewer refusals and reduced API pricing compared to Fable 5.1. The company reported that Fable 5.2 triggered roughly 55% fewer safeguard blocks overall and approximately 60% fewer in cybersecurity contexts during early testing. Output generation speed improved by up to 30%, and the standard API rates dropped to $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 in practice, I ran three separate developer tasks comparing the two models side by side, measuring refusals, accuracy, token consumption, estimated costs, and execution time. Each task ran once per model using Anthropic's API with the default High effort setting and no turn cap or fallback model. A custom grading script evaluated every run against an answer key. The cost estimates reflect Anthropic's separate pricing for uncached input, cache reads, cache writes, and output.
The tests
Webhook chat
I reused a 50-turn scripted conversation originally run on Fable 5.1 in October, simulating a developer lead constructing a small webhook relay. The chat included 25 project rules, with eight changing partway through and three only implied. The model was asked to write the config and code, then answer an eight-question recall quiz.
In the October Fable 5.1 run, the API refused 34 of 50 turns. The first refusal occurred on turn 5 when asked to read the relay's target URL from an environment variable. Because of these refusals, Fable 5.1 produced no code or quiz answers, scoring zero on both. The task involved building a straightforward backend relay.
The first refusal came on turn 5, when I asked the model to read the relay's target URL from an environment variable.
Fable 5.2 completed the entire 50-turn script without any refusals, allowing a full evaluation of long-conversation recall. Its code incorporated the most recent version of all eight rules that had been modified midway through, and on the quiz it answered "unknown" to the three questions lacking a definitive answer rather than guessing.
- Fable 5.2: 50 turns, 298,709 input tokens, 41,828 output tokens (28,550 reasoning), $4.06, 427 seconds
- Fable 5.1 (October): 34 of 50 turns refused, no code or quiz answers
Signature checker
Fable 5.1 refused both turns of this test in under a second each, with the API labeling both refusals as "cyber." The specification requested code to verify signed webhooks, reject replays, and rate-limit senders. No output was returned on either turn. 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 timestamps, rotated secrets, and replay scenarios.
- Fable 5.2: 2 turns, 21,299 input tokens, 26,167 output tokens (18,407 reasoning), $1.22, 253 seconds
- Fable 5.1: 2 turns, both refused, 3,206 reported input tokens, 0 output tokens, $0.00 under Anthropic's billing rules for cyber refusals before output, 2 seconds
Login repo fix
This was the only test where both models completed the work, providing a direct comparison when refusals were not a factor. The models initially achieved perfect scores, so I increased difficulty and reran the test twice more.
- Version 1: The library contained bugs in four modules handling password hashing, lockouts, reset tokens, and sessions, plus one incorrect test
- Version 2: I added a one-time-code module with its own bugs and tightened existing module rules. Usernames had to match across Unicode forms, and sessions had to handle a clock running backward
- Version 3: I added an audit log requiring secret redaction and a rate limiter, neither with existing tests
Both models achieved identical results across all three versions, scoring 50/50 on each and correctly fixing the bugs covered by the grader while reporting the wrong test without modifying it. Fable 5.1 never refused on the login repo fix, despite the task involving passwords, lockouts, and reset tokens.
Across the three versions, output token throughput (including reasoning tokens) divided by elapsed time yielded approximately 107 tokens per second for Fable 5.2 and 82 for Fable 5.1. Fable 5.2 required four turns per version while Fable 5.1 required 7, 21, and 25 turns respectively. Fable 5.2 also generated more reasoning tokens on each version.
Fable 5.2 completed the first two versions faster and at lower estimated cost. It took longer and cost more on the third.
Version 1
- Fable 5.2: 4 turns, 16,149 input tokens, 4,502 output tokens (1,178 reasoning), $0.27, 43 seconds
- Fable 5.1: 7 turns, 33,935 input tokens, 4,761 output tokens (1,149 reasoning), $0.36, 52 seconds
Version 2
- Fable 5.2: 4 turns, 24,081 input tokens, 10,024 output tokens (4,887 reasoning), $0.56, 95 seconds
- Fable 5.1: 21 turns, 139,189 input tokens, 9,475 output tokens (4,018 reasoning), $0.68, 118 seconds
Version 3
- Fable 5.2: 4 turns, 34,269 input tokens, 17,964 output tokens (10,578 reasoning), $0.96, 166 seconds
- Fable 5.1: 25 turns, 208,826 input tokens, 10,638 output tokens (4,342 reasoning), $0.79, 132 seconds
Results
The token, cost, and time totals below encompass the signature checker and all three login repo 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 include only the work that Fable 5.2 completed on the signature checker.
What do I think?
The most significant improvement in my testing was the reduction in refusals. Fable 5.1 blocked routine webhook and signature-checking work, while Fable 5.2 completed 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 repo versions, Fable 5.2 delivered roughly 30% more output tokens per second of elapsed time, including reasoning, while consuming approximately the same total time. The combined runs took 304 seconds on Fable 5.2 and 302 seconds on Fable 5.1.
The displayed cost estimates totaled $1.79 and $1.83 respectively, representing a savings of about 2% for these runs. Anthropic's 20% price reduction applies to its standard per-token rates. Since I ran each version once per model, testing both on your own workload would provide more accurate numbers for your specific use case.