Anthropic has conducted a series of major announcements, including the release of the high-performance and affordable Claude Sonnet 5 model, the restoration of access to flagship Fable 5 and Mythos 5 models for the global market, and the launch of a specialized platform for scientific research.
What Happened
Anthropic introduced Claude Sonnet 5, whose performance is comparable to Claude Opus 4.8, while reducing usage costs by 60% to $3 per 1 million input tokens. Following the lifting of US export restrictions, the company restored global access to the Fable 5 and Mythos 5 models. Additionally, a research platform for scientists was launched, and hidden Unicode characters were discovered in the system prompts of the Claude Code tool, designed to detect model distillation attempts by Chinese developers.
Context
Previously, access to Anthropic's most powerful models was limited by US export rules, which hindered the global use of the company's technologies. Simultaneously with increasing accessibility, the company needs to protect its intellectual property from unauthorized copying and weight distillation via third-party APIs.
Why It Matters for the Industry
The release of Sonnet 5 significantly intensifies competition with OpenAI in the segment of economical yet intelligent models, forcing the market to rethink pricing. The use of steganographic protection methods (hidden markers) in Claude Code sets a new industry precedent for protecting proprietary models from distillation.
Why It Matters for Users
Developers gain the ability to build complex AI agents with a substantially lower Total Cost of Ownership (TCO) thanks to Claude Sonnet 5. Scientists receive a specialized environment for integrating LLMs into their research workflows, and global users regain access to top-tier Fable and Mythos models.
What Is Not Yet Known / Limitations
The use of hidden characters to protect models may create difficulties when auditing system transparency and could hinder researchers striving for full openness of inference mechanisms.
Sources
Author
Look at AI, Editorial Team