Anthropic launches new Claude AI models to reduce operating costs, improve performance and allow longer running AI applications AI: The rise of intelligent code writers By Rob Price , on August 8, 2024 110 SHARES Anthropic today launches its new artificial intelligence models – Claude Fable 5.1 and Claude Mythos 5.1 – designed to help users with coding, knowledge work, scientific research, and other demanding AI applications. With its latest AI models, Anthropic is attempting to address the three major trends it’s seeing in the AI-model market: lower operating costs, increased model performance and more accurate safety limitations. The company aims for Fable 5.1 to make fewer unnecessary interventions and also handle more sophisticated tasks.
The release comes at a time when AI model makers are increasingly jockeying not just to build the most intelligent models, but to also offer affordable, secure and highly capable AI systems.
Claude Fable 5.1: What is new?
The widely-available version is Fable 5.1, pitched to those working in coding, research, and other forms of knowledge-based work. In addition, this model also becomes a much more efficient choice when using previously provided contexts or prompts over and over again. In particular, AI token rates have been altered.
“By nearly cutting edge read prices by 75%, to $0.25 / million tokens,” says Anthropic on its blog.
Standard input- and output tokens remain at the higher $10/ million tokens, but lower cache reading rates make these prices much cheaper for high performance, recurrent workloads. The company estimates costs could be slashed by about 25% for some jobs and over 45% for agent workloads “that require significant context reuse.” Fewer false-positive AI restrictions One other benefit touted by Anthropic for Fable 5.1 is improved performance regarding AI safety measures.
According to the company, the new version is built to avoid triggering safety stops that are either unnecessary or mistakenly identified. “We reduced unjustified AI safety-stops-less likely to over-block harmless uses when properly deployed,” Anthropic writes on its company blog. But while this does lower some of the roadblocks when dealing with AI security and code-based functions that could have previously caused trouble-such as researching cybersecurity or creating code related to illegal hacking operations that are ultimately intended for research-the system does still have safety nets.
These safeguards are tweaked based on the kind of prompt that comes in, and where the AI is working.
The balance that Anthropic wants to strike is on using its AI to handle software engineering, or to assist with science and technology, without the system acting too overly cautious of perfectly harmless requests. Mythos 5.1 is aimed at specialised users In addition to Fable 5.1, there’s Mythos 5.1. This particular model from Anthropic is designed for special-use cases. Fable 5.1 is more general, whereas Mythos 5.1 requires customers to be vetted by Anthropic to access this model.
Areas of application include advanced AI
“Mythos provides more tailored safety guardrails to approved partners leveraging our technology for complex security, or for other use cases requiring specialized technical research or the use of a particular model configuration. Interestingly, both models use the exact same architecture as one another, differing primarily in their safeguard set up and customer access.” writes Anthropic on its blog. It notes that mythos was able to achieve superior results on certain benchmarks under low safety limitations.
“In particular for coding and science benchmarks in internal evaluations, Mythos achieved 5-50% more completions on technical benchmarks with adjusted safeguards” Anthropic wrote. Stronger performance in coding and researchAnthropic says that their latest models boast more sophisticated skills. In addition to performing tests across coding and research based tasks (using its Terminal-Bench test in addition to Humanity’s Last Exam) , Anthropic highlights some use cases for its new AI models including: Using Fable to help optimize GPU performance Working with Fable to create a high-resolution map of Venus based on existing orbital imagery Developers of software also benefit from higher productivity and the ability to run longer, more intricate workflow.
New privacy-focused option For Businesses
“Anthropic is launching Enterprise Frontier Safeguards a more privacy-focused, infrastructure-based offering that will provide customers greater control over their data and monitoring and provide infrastructure that customers bring in house,” the company states in its blog, adding “We assure customers that enterprise customer data has never been shared without explicit permission for training our foundation models.”
What Anthropic Fable 5.1 launch means for the AI industry
The trends highlighted by Anthropic are indicative of the direction the whole AI model market is pushing toward. While accuracy, intelligence, and ability are always the priority, those considerations come packaged alongside a wider set of issues: costs, privacy, security, and functionality for long-form tasks are of significant importance, too. As more businesses look to secure enterprise AI, Anthropic’s approach to market is trying to meet those specific needs. Between Fable and Mythos, the company seems intent on creating tiers for particular AI uses, pushing both its general-purpose commercial models and its highly specialised options through similar developments in speed, price, and safety.
With Anthropic also joining Google and OpenAI in a privacy-focused approach, the competition in the field is likely to only intensify, as the pursuit for both advanced and affordable AI continues.
