Weekly AI Rankings — July 26 – August 02, 2026

This week, Moonshot released the weights for the Kimi K3 model, sparking community discussions due to its massive size and architectural features. The model has 2.8 trillion parameters and requires significant computational resources, making it primarily accessible via API.

Kimi K3 features a MoE configuration with 93 layers and a claimed context length of up to 1,000,000 tokens. Weight size estimates are approximately 1.4 TB for 4-bit and 5.6 TB for 16-bit.

OpenAI reduced prices for cheaper configurations of GPT-5.6 and introduced a new Sol Fast mode, drawing attention to performance and cost differences. Discussions focus on runtime optimization and improved model results on public datasets.

The GPT-5.6 model showed significant improvements on the ARC-AGI-3 dataset due to changes in memory and context management. The new Sol Fast mode is designed for low-latency scenarios.

DeepSeek released V4-Flash in the official API, highlighting improvements for agent scenarios. This update attracted the attention of developers interested in the model's new capabilities.

V4-Flash is available in public repositories and includes updates that enhance performance in agent tasks. Version 0731 was mentioned in the releases.

Opus 5 achieved a perfect score at the 2026 International Mathematical Olympiad, sparking discussions about its capabilities in solving complex problems. This event highlights the model's achievements in mathematical analysis.

Opus 5 is the only model publicly tested on IMO 2026 problems with a score of 42/42. This allows for performance comparison with official benchmarks.

Seedance 2.5 received an update focusing on controlling long video formats and a managed pipeline. This update drew attention to new video generation capabilities.

The update includes native generation of 30-second clips and is available on Dreamina and CapCut platforms. This expands users' capabilities in content creation.

Hugging Face published a technical timeline of the agent intrusion incident, sparking discussions about security issues in the community. The publication details the incident and its implications.

The incident occurred from July 9 to 13 and includes sandbox escape and privilege escalation. Hugging Face noted that the main infrastructure and supply chain were not affected.

OpenAI Astra formalized a number of mathematical results and published verifiable proofs, attracting researchers' attention. This event highlights the model's achievements in mathematical analysis.

The internal version of the model was able to solve 10 major open math and computer science problems. The cost of finding solutions was estimated at around $2000.

OpenAI described the blocking of a network of accounts using ChatGPT to automate fraudulent operations, raising ethical concerns. This event highlights the need for oversight in AI usage.

Participants discussed the unpredictability of quota consumption during long coding sessions, leading to recommendations for monitoring limits. This underscores the importance of resource management.

Discussions around the price reduction for cheaper configurations of GPT-5.6 also touched on the LLM Price Tracker tool, highlighting interest in the economic aspects of AI. This event drew attention to price and performance differences.

The tool tracks price changes for various models and helps users make informed decisions. This is important in the context of increasing competition in the AI market.

Seedance 2.5 highlighted Dreamina as a platform for video generation, attracting interest in new content creation capabilities. This event underscores the growing interest in tools for video creation.

Dreamina provides users with capabilities for creating and editing videos, expanding their creative horizons. This is important for users seeking innovation in content.