Anthropic released its new large language model (LLM), Claude 3, this morning that appears to be the world’s most powerful LLM thus far; surpassing OpenAI GPT-4 and Google Gemini Advanced on common benchmark tests; while Amazon already integrated these models into Bedrock — their cloud platform for developing and running AI services — in just 24 hours after Anthropic released it.
Anthropic unveiled today three new Claude 3 models: Opus, Sonnet and Haiku in order of decreasing intelligence. Of note is the fact that they were trained using synthetic data generated by AI rather than by human authors directly, hopefully mitigating concerns of model collapse.
Bedrock AI’s fully managed service, launched early 2023 as a single API through which customers could access multiple models, will now include access to its middle-tier model Claude 3 Sonnet with Opus and Haiku expected “soon.”
Amazon Bedrock Pricing Model For Claude 3 Sonnet
Amazon provides several pricing models to customers of its Bedrock service, not only for Claude 3 Sonnet but for other foundation LLMs available through this managed service. Here is a snapshot from Amazon showing its pricing models directly related to Claude 3 Sonnet:
Claude 3 Sonnet is more expensive than its older, smaller and less powerful (but less compute resource demanding) counterpart Claude Instant, while it is cheaper per 1,000 tokens than Anthropic’s previous flagship model, Claude 2.
Access to Claude 3 Sonnet can also be purchased hourly; it is more costly than both Instant and 2 to run, however.
Comparative to other foundation LLMs available on Bedrock, Claude 3 Sonnet stands out in terms of pricing. It ranks among the more costly models on this platform.
Why the addition of Claude 3 on Bedrock matters This episode’s presence on Bedrock is noteworthy since Amazon last year announced a $4 billion investment in Anthropic, which is under investigation by the U.S. Federal Trade Commission as anticompetitive practices may threaten Anthropic’s survival in AI industry.
Yet the Amazon and Anthropic partnership is by no means exclusive; customers can access Claude 3 directly on Anthropic’s website outside of Amazon; additionally, Amazon provides access to numerous LLMs from Bedrock offered by AI21 Labs, Cohere Meta Mistral Stability AI as well as themselves.
Amazon only announced two weeks ago the addition of Mistral’s open-source 7B and Mixtral 8x7B models from French startup Mistral to its Bedrock cloud services, only for Mistral to then announce a brand new, closed model called Mistral Large along with partnership and investment from Microsoft that will limit it solely to Microsoft Azure and Mistral’s website. These rapid developments show just how competitive this market for attaching cutting-edge AI models and APIs to cloud services has become.
Dr. Swami Sivasubramanian, Vice President of Data and AI at AWS, expressed great excitement for Anthropic’s collaboration with AWS customers today in a blog post, writing:
“Our customers and partners continue to be impressed by the innovative applications they can create with Claude on Amazon Bedrock, as well as its unrivalled capability of rapidly, securely, and responsibly deploying generative AI applications utilizing knowledge bases, guardrails and model evaluation. Thanks to its ease-of-use, many leading startups, enterprise businesses, and government organizations have chosen Amazon Bedrock’s managed service for deploying their generative AI applications – and we expect this trend to accelerate further with today’s news.”
AWS’s announcement today is part of its broader strategy for leadership in Gen AI. By investing across all layers of the stack–infrastructure, models, and user-facing apps–AWS strives to make AI easier for customers to leverage more efficiently, quickly, and comprehensively, thus speeding innovation and providing new experiences to end users.
Amazon reports that more than 10,000 organizations worldwide are currently using Amazon Bedrock to explore and deploy generative AI applications.
As expected, OpenAI may unveil its own answer to Claude 3, GPT-5, in the near future – possibly as early as today! We’ll keep you up-to-date.
