Weekly AI Rankings — June 07 – June 14, 2026

This week, Anthropic suspended access to Fable 5 and Mythos 5 at the request of U.S. authorities, sparking discussions about geopolitical risks in the AI sector. This decision also affected user and developer access to these models.

Fable 5 was introduced as a public version of Mythos with additional safeguards and data handling rules, priced at $10/$50 per million tokens.

This week, Anthropic suspended access to Fable 5 and Mythos 5 at the request of U.S. authorities, sparking discussions about geopolitical risks in the AI sector. This decision also affected user and developer access to these models.

Mythos 5 retains the capabilities of Fable 5 but includes additional safety measures, making it more suitable for public use.

This week, Anthropic rolled back its policy of hidden degradation of Claude's responses, returning to explicit refusals for certain queries. This decision was made following criticism from users.

Claude remains a popular model, and its behavior has become more predictable for developers following the policy changes.

This week, Moonshot AI released Kimi K2.7 Code, which improves coding quality and requires 30% fewer tokens compared to the previous version. This highlights the growing pressure of open-weight models on closed solutions.

Kimi K2.7 Code offers enhanced capabilities for developers, making it competitive in the AI tools market.

This week, MiniMax introduced M3 as a major open-weight release with 428B parameters and a unique MiniMax Sparse Attention scheme. This event drew attention to the efficiency of long-context architectures in models.

M3 is a multimodal model that supports various data types and tasks, making it appealing to researchers and developers.

This week, Google released DiffusionGemma, a diffusion LLM, emphasizing its high generation speed. Discussions focused on its capabilities for working with text and images.

DiffusionGemma has 26B parameters and 4B active, allowing it to generate text faster than many similar models.

This week, OpenAI added the ability for manual limit resets for Codex, enhancing its functionality during peak loads. This change was discussed in the context of comparisons with other tools.

Codex now allows users to generate and deploy websites from text descriptions, significantly speeding up the prototyping process.

This week, SemiAnalysis assessed the real cost of ChatGPT and Claude subscriptions, sparking discussions about their economic viability for heavy users. This highlights the importance of cost analysis in AI.

The analysis showed that subscriptions may not be cost-effective for heavy usage, raising questions about their attractiveness for professionals.

This week, CacheLib received an update after a long hiatus, drawing attention to its hybrid caching capabilities. This update may help optimize memory costs in AI infrastructure.

CacheLib provides a solution for optimizing memory usage in AI applications, making it a useful tool for developers.