SI Agent News
DevelopingClaudeSecurity

Anthropic releases Claude Fable 5.2 with fewer refusals and lower costs

The new model update aims to reduce false positives in security filters while offering a 20% price cut on standard API rates.

20%reduction in standard API rates

The short version

Why it matters

For developers, this means fewer interruptions from the AI's safety filters when working on legitimate security code. It also provides a more predictable experience for long-conversation tasks where the model must remember specific project rules.

Reducing safety filter friction

Anthropic recently released Claude Fable 5.2, a model update designed to improve performance and reduce instances where the AI refuses to complete a request. In early testing, the company reported that Fable 5.2 triggered about 55% fewer safeguard blocks than the previous Fable 5.1 version. This improvement is particularly notable in cybersecurity contexts, where the company saw a 60% reduction in blocks.

These safeguards are intended to prevent the model from assisting in malicious activities, but they can sometimes trigger on benign work. For example, a developer might be writing code to verify signed webhooks or manage rate limits, which can look like cyberattacks to an automated classifier. Fable 5.2 aims to distinguish between these legitimate developer needs and actual threats more accurately.

In a head-to-head comparison, Fable 5.1 refused both turns of a test involving a signature checker because the API labeled the request as a cyber risk. In contrast, Fable 5.2 handled the same specification and passed all 50 hidden tests. This suggests that the new model is better at understanding the context of security-related programming tasks without triggering a refusal.

Speed and reasoning improvements

Beyond safety improvements, Fable 5.2 is designed to be faster and more efficient at generating content. Anthropic states that the new model generates output up to 30% faster than Fable 5.1. This speed increase is important for developers who rely on real-time interactions and want to see code generated quickly as they work through complex logic.

The model also shows different behavior regarding reasoning tokens, which are the internal steps a model takes to think through a problem. In a series of login library fixes, Fable 5.2 produced more reasoning tokens than Fable 5.1 while maintaining similar total elapsed times. This indicates that the model may be performing more internal processing to arrive at the correct solution for difficult bugs.

When comparing the two models on a 50-turn scripted chat, Fable 5.2 finished the entire script without a single refusal. It successfully recalled project rules that had changed midway through the conversation and correctly identified which questions had no answers. This level of consistency is a key goal for the Fable 5.2 update, ensuring that the model remains helpful over long interactions.

New pricing and cost dynamics

Anthropic has implemented a 20% price cut on the standard API rates for Fable 5.2. The new rates are set at $8 per million uncached input tokens and $40 per million output tokens. These prices are lower than the $10 and $50 rates previously associated with Fable 5.1, making it a more affordable option for high-volume users.

However, the actual cost savings in a real-world workflow can vary depending on the specific task. In a test involving three different versions of a login library fix, the total estimated costs for both models were very similar, showing only a 2% difference. This suggests that while the per-token price is lower, the model's increased reasoning and different turn counts can balance out the savings.

The cost of a task is determined by several factors, including uncached input, cache reads, cache writes, and output. Because Fable 5.2 often completes tasks in fewer turns than Fable 5.1, it can sometimes be cheaper for simple tasks. For more complex tasks that require more reasoning, the higher number of reasoning tokens might offset the lower base price.

Performance in complex coding

When both models are able to complete a task without refusing, their performance becomes a matter of efficiency and accuracy. In the login library tests, both Fable 5.1 and Fable 5.2 achieved perfect scores on three different versions of the library. They both successfully fixed bugs related to password hashing, lockouts, and session management while correctly identifying a wrong test.

Fable 5.2 demonstrated a higher output rate of about 107 tokens per second compared to 82 tokens per second for Fable 5.1. This means that while the total time to complete a task might be similar, Fable 5.2 is processing more information in the same amount of time. This efficiency is a primary goal for the Fable 5.2 update as it scales for larger workloads.

The comparison shows that Fable 5.2 is particularly effective at handling multi-turn tasks with changing rules. In the webhook chat test, it correctly used the most recent version of eight rules that changed during the conversation. This ability to maintain context over long sequences is vital for developers who need the AI to act as an agent on a large, evolving codebase.

What to watch for next

The transition to Fable 5.2 marks a shift toward making AI models more permissive for legitimate technical work. Developers should look for how these changes affect the

try_it-placeholder-logic-not-applicable-here-but-following-rules

What to watch

Try it

Check your current coding prompts to see if they frequently trigger 'cyber' refusals and test them on the Fable 5.2 model to see if they pass.

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 Wire

Listen to our daily AI shows every morning: on your commute, in the shower, wherever you are

Listen to the daily shows

First reported by The New Stack · 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.

More AI news