Anthropic Unveils 'Claude Fable 5.1' and 'Claude Mythos 5.1,' Bolstering Agent Performance and Safety Guardrails
TECHWORLD ·
✦ AI Summary
Anthropic unveiled its next-generation AI models, Claude Fable 5.1 and Claude Mythos 5.1.
Fable 5.1 will be generally available, while Mythos 5.1 will be released in a limited offering through a trust-based access program.
Fable 5.1 has improved coding and agent task performance, and cache read pricing was cut by 75%.
Anthropic on the 1st (local time) unveiled its next-generation AI models, "Claude Fable 5.1" and "Claude Mythos 5.1." The two models are focused on improving complex agent tasks and coding performance.
It also outlined how they will be offered. Fable 5.1 will be generally available, while Mythos 5.1 will be released in a limited offering through a trust-based access program. Along with that, the company is moving to ease costs for enterprise customers, strengthen data control, and reinforce safety guardrails in the biological and cyber domains.
In particular, Fable 5.1 is an upgraded version of the existing Fable 5. The model has improved performance on complex multi-step agent tasks, including coding and knowledge-based work.
Anthropic said it cut cache read pricing by 75% to reduce costs for long-running agent tasks and complex coding processes.
On pricing, Fable 5.1's cache read feature was introduced. Cache read is a feature for recalling previously processed and stored inputs. The rest of Fable 5.1's pricing remains the same as Fable 5: USD 10 per 1 million input tokens and USD 50 per 1 million output tokens.
On performance, enterprise pretests showed improvements in coding and agent task performance. Sualeh Asif, machine learning director at SpaceXAI, said Fable 5.1 has strong task verification capabilities and can carry out difficult coding tasks from start to finish.
Sualeh Asif also said Fable 5.1 recorded the best performance among the models tested in SpaceXAI's coding benchmark, "CursorBench 3.2." Fable 5.1 scored 73.4% at the highest reasoning level.
In the browser-based advanced browser agent evaluation, Fable 5.1 posted a task completion rate of 82%. Opus 5 recorded 74%, and Fable 5 recorded 57%. Fable 5.1 used fewer tokens than Opus 5 and Fable 5.
The performance gains were presented as improvements in long-running task execution. In the MongoDB evaluation, Fable 5.1 was described as analyzing service code and documentation, generating a new design, implementing a prototype over several hours without additional intervention, and then verifying the implementation results.
At the same time, Anthropic said it reduced cases in which safety guardrails block normal requests while improving model performance. The latest biological safety guardrails recorded an 85% reduction in the frequency of interventions for normal requests compared with the guardrails at the time of Fable 5's launch, and Claude Code's cyber safety guardrails were expected to reduce the average frequency of interventions per session by about 60% compared with Fable 5.
These adjustments were presented as being aimed at supporting the normal security work of cybersecurity personnel. Vulnerability scanning was cited as an example of normal security work, along with support for using Fable 5.1.
The introduction of data control functions for enterprise customers was added to this. Anthropic reflected customer feedback on Fable 5's data retention policy and worked with more than 100 customers in finance, healthcare, manufacturing, telecommunications, legal, retail, and the public sector on developing Enterprise Frontier Safeguards (EFS).
EFS is a frontier model misuse detection method, and related data are stored in customer-controlled cloud infrastructure rather than by Anthropic. In this environment, customers can apply their own encryption keys, access policies, and audit logs, and activity data is stored in Amazon S3, Azure Blob Storage, and Google Cloud Storage.
An automated system analyzes traffic over a set period to detect signs of serious misuse across multiple sessions and accounts. If a problem is detected, the relevant signals are sent to the customer, and the subsequent decision is handled by internal personnel at the customer, without direct review by Anthropic staff.
EFS is set to be rolled out gradually starting this fall across Claude Code, Claude Enterprise, Claude Platform, Amazon Bedrock, Claude Platform based on AWS, Google Agent Platform, and Microsoft Foundry. Before EFS is adopted, Zero Data Retention (ZDR) for Fable 5 and Fable 5.1 will be provided to eligible customers.
Along with the simultaneous release of Mythos 5.1, Anthropic decided not to make Mythos 5.1 generally available and set its operating model to a limited rollout through Anthropic's trust-based access program. It also said separate safety guardrails were applied to Mythos 5.1, and that the purpose of those guardrails is to support research in cybersecurity and the life sciences.
Anthropic said it conducted tests across a variety of scientific research fields for Fable 5.1 and Mythos 5.1, and has drawn up plans to verify the two models' research usefulness based on early testing. It is also pursuing plans to expand AI's role in the process of scientific discovery.
Source: TECHWORLD · Kim Seung-gi
Original: https://www.epnc.co.kr/news/articleView.html?idxno=406374
References
This article was produced with the help of an automated content generation algorithm.
Source: TECHWORLD
View originalThis article was summarized and organized by BizCrush based on the original article from TECHWORLD. For exact quotations and full details, please refer to the original article.