Anthropic updates Claude Fable to reduce coding refusals
The new Fable 5.2 model aims to complete more security-related coding tasks while lowering API costs and increasing output speed.
The short version
- Fable 5.2 triggered about 55% fewer safeguard blocks overall than Fable 5.1 in early testing.
- The update reduced cybersecurity-related refusals by approximately 60% across several platforms.
- Anthropic lowered standard API rates by 20% for the new model version.
- Testing showed Fable 5.2 can generate output up to 30% faster than its predecessor.
Why it matters
For developers, this means fewer interruptions when working on legitimate security features like webhooks or signature verification. It also offers a more predictable cost structure for high-volume coding tasks.
Reducing false positives
Anthropic recently released Claude Fable 5.2, a model designed to improve performance on coding tasks. A primary goal of this update is to reduce the number of times the model refuses to complete a request due to its internal safety filters. In early testing, the company reported that Fable 5.2 triggered about 55% fewer safeguard blocks than the previous version, Fable 5.1.
These refusals often occur when a model misidentifies benign work as a security threat. For example, developers building legitimate security features like webhooks or signature checkers might find their requests blocked by a model's cyber classifier. Anthropic states that the new version is intended to be more discerning in these scenarios.
In specific cybersecurity testing, the reduction in refusals was even more pronounced. The company noted that Fable 5.2 triggered about 60% fewer blocks in cybersecurity tasks across various platforms, including claude.ai and Claude Code. This suggests a refinement in how the model distinguishes between harmful intent and standard development work.
Performance in coding tasks
Beyond safety improvements, Fable 5.2 introduces several performance optimizations for developers. Anthropic claims the new model generates output up to 30% faster than Fable 5.1. These improvements are intended to streamline workflows where speed and efficiency are critical for iterative coding.
In head-to-head comparisons, the model demonstrated its ability to handle complex instructions. During a webhook chat test, Fable 5.2 completed a 50-turn script without a single refusal. It successfully followed project rules and answered a recall quiz by identifying which questions had no available information rather than guessing.
The model also showed strong results on a signature checker task. It successfully passed 50 hidden tests, including difficult edge cases involving rotated secrets and timestamps. While Fable 5.1 refused both turns of that specific test, Fable 5.2 completed the work in 253 seconds with an estimated cost of $1.22.
Cost and token efficiency
Anthropic has also adjusted the pricing structure for the new model to make it more accessible for high-volume use. The standard API rates for Fable 5.2 are $8 per million uncached input tokens and $40 per million output tokens. These rates represent a 20% decrease from the $10 and $50 prices associated with Fable 5.1.
While the price cut applies to standard rates, real-world costs can vary based on the complexity of the task. In a login library fix test, both models completed the work with similar total costs. However, Fable 5.2 often reached the solution in fewer turns, which can impact the overall efficiency of the interaction.
The model also showed a different behavior regarding reasoning tokens. In some tests, Fable 5.2 produced more reasoning tokens per version than its predecessor. Despite these differences in internal processing, the total elapsed time for several tasks remained very close between the two models, with Fable 5.2 showing a slight edge in output tokens per second.
Comparison of model behaviors
Testing the two models side-by-side reveals distinct differences in how they handle complex logic. For instance, in a multi-version login repository fix, Fable 5.2 consistently completed all tasks in four turns. In contrast, Fable 5.1 required significantly more turns—ranging from 7 to 25—to complete the same set of bugs.
This difference suggests that Fable 5.2 may be more direct in its problem-solving approach. By requiring fewer turns to reach the correct answer, the model can potentially reduce the total number of tokens processed for a single task. This is particularly useful for developers who need to maintain a high level of accuracy while managing API usage limits.
The results indicate that while Fable 5.1 was reliable for simple tasks, it struggled with the specific security-related prompts that Fable 5.2 handled smoothly. The new model appears to strike a better balance between maintaining safety guardrails and allowing developers to perform necessary security-related programming. These findings highlight the ongoing effort to make AI tools more useful for specialized engineering roles.
What to watch
- How the 20% price cut affects high-volume enterprise users.
- Whether the reduction in refusals leads to any unintended safety trade-offs.
Try it
If you use Anthropic's API for security-related coding, try running a signature verification task on Fable 5.2 to see if it completes without a refusal.
Sources and links
Liked it? Say why
Add a line for other readers.
No account needed. We never show your email. Comments with links, ads or abuse are held and not shown. Your comment shows after a quick check.
Listen to the AI Wire headlines
Listen to the AI WireListen to our daily AI shows every morning: on your commute, in the shower, wherever you are
Listen to the daily showsFirst reported by thenewstack.io · Read the original. Written by the SI Agent News Desk from the reporting above.Written with AI from the source above. Every number is checked against the source. Spot a mistake? Corrections.
