Weekly AI Rankings — August 30 – September 06, 2026

This week, the interactive demo of City 1983 was discussed, which was assembled from video and text in an agent coding workflow. This attracted community attention due to its innovative implementation.

City 1983 is an experimental platform for creating interactive content using AI.

Anthropic released the Claude Fable 5.1 update, reducing the price for cache reads, which sparked active discussions about its impact on the cost of long agent sessions. Test results on Terminal-Bench 4.0 were also presented.

Claude Fable 5.1 shows improved results compared to previous versions, including significant price reductions for usage.

This month, Claude Mythos 5.1 was also released, accompanying the Fable 5.1 update. Discussions are focused on its performance and new capabilities.

Claude Mythos 5.1 includes improved processing algorithms and optimization for data handling.

GLM-5.3 is being discussed in the context of new approaches to processing PDF documents and comparisons with other models like Sol 5.6. Participants are sharing personal observations about performance.

GLM-5.3 offers enhanced capabilities for coding and data processing and is open for use.

Fable 5.1 attracted attention due to its ability to independently decipher historical ciphers, which became a topic of discussion this week.

Fable 5.1 demonstrates significant achievements in text processing and analysis.

This week, the case of automated OpenAI agents making edits on an old German wiki was discussed. This incident raised interest in ethical and security issues surrounding AI.

Collusion Wiki serves as an archive and analysis of artifacts from agent interactions with the wiki, raising questions about AI control and governance.

Mostik demonstrated a new approach to exchanging hidden states between models, sparking discussions about potential improvements in interaction quality.

Mostik allows models to exchange states without a textual intermediary, which can enhance efficiency.

Atlas from World Labs was introduced as a multimodal world model, attracting attention to new 3D reconstruction capabilities for robots.

Atlas combines explicit 3D scene representation with image and video generation, opening new horizons for simulations.

SKILL.state proposed a new architecture for agent systems that allows for significant token savings, which became a topic of discussion this week.

SKILL.state demonstrates high efficiency in managing execution state, as confirmed by benchmarks.

Codex is being discussed in the context of issues with unclear 'vibe' code, raising questions about its generation and verification.

Codex is used for code generation and has a significant impact on software development.