Weekly AI Rankings — August 16 – August 23, 2026

This week, the anonymous model Ox Alpha was discussed after its release on OpenRouter, featuring a context window of 1 million tokens. Early tests indicated strong performance on a small sample, but broader evaluations revealed a drop in effectiveness.

Ox Alpha features a context window of approximately 1 million tokens and is available via OpenRouter. Independent tests revealed a success rate of 59% on the full DeepSWE benchmark.

Microsoft announced MAI-Image-2.5-Pro, which now ranks among the top in text-to-image generation. Discussions focus on the model's pricing structure and its comparison with competitors.

MAI-Image-2.5-Pro is priced at approximately $108.5 for 1000 images, with quality differences depending on prompts. The model is comparable to leaders in the field.

OpenAI announced a three-month price reduction for GPT-5.6 Sol, drawing user attention to the new rates. Discussions revolve around the impact on the cost of long sessions.

The new rates for GPT-5.6 Sol are $4/$20 per million tokens for short contexts and $8/$30 for long ones. This price reduction could significantly lower costs for users.

DeepSeek announced V4-Flash-Vision-Exp, aimed at multimodal agent scenarios, which sparked interest in its capabilities. Discussions focus on support for vision functions and new features.

The model supports image tokenization and Files API for batch processing. Documentation for V4-Flash-Vision-Exp has already been published.

This week, a custom skill for Claude Opus 5 was discussed, which enhances the model's output verification. This drew attention to issues of stability and repeatability of results.

The case using the skill layer showed improved stability on ARC AGI 3, although detailed metrics were not provided. This highlights the importance of formal output verification.

DeepSeek Harness became the fastest-growing AI repository this week, attracting over 140,000 stars. Discussions focus on its functionality and lack of documentation.

The project includes a CLI and local Web UI, requiring environment setup for launch. Users note interest in extending its functionality.

The desktop-fly repository featuring 3D visualization of a fly based on the FlyWire map attracted attention due to its unique visualization. Discussions revolve around its application in biological research.

The project utilizes a map with approximately 139,000 neurons of the fruit fly and provides source code for visualization. This is useful for researchers in the field of neurobiology.

DeepSeek Harness Desktop also gained popularity this week due to its functionality and rapid growth. Discussions focus on its applications and capabilities.

The Desktop version provides users with access to the functionality of DeepSeek Harness in a more convenient format, facilitating work with the project.

Microsoft AI Playground is discussed in the context of the MAI-Image-2.5-Pro announcement, drawing attention to its capabilities in image generation. Discussions focus on quality and pricing.

MAI-Image-2.5-Pro offers competitive pricing and high quality, making it attractive to users. This underscores the importance of the AI Playground platform.

HarnessOpt-Bench was released by Scale AI to evaluate harness rewriting, sparking interest in its methodology. Discussions focus on its applications and objectives.

The benchmark allows models to test whether they can improve or rewrite another agent's harness, opening new opportunities for optimization.