Top 5 AI Models of the Week
August 16 – August 23, 2026
This week, there is a growing interest in models with new announcements and price changes, impacting their popularity and usage.
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.
Rankings are based on independent discussion volume and engagement, with no manual position editing.
About methodology
Algorithm analyzes thousands of mentions in technical channels, counting unique sources and discussion quality. Positions are determined automatically based on popularity and relevance for the Russian-speaking AI community.
Updated: August 23, 2026 18:00 MSK
Frequently asked questions
How is the weekly AI models ranking compiled?
The algorithm aggregates model mentions across 50+ Russian-language AI channels and chats over a week, counting independent discussions, unique sources, and engagement (reactions, replies). Positions are determined automatically — no manual editing.
How does this differ from Hugging Face or LMArena?
HF and LMArena rank models by technical benchmarks and user votes. This ranking measures real practitioner discussion volume in the Russian-speaking engineering community — what people are actually trying, comparing, and debating this week.
Can I see archive of previous weeks?
Yes, the archive navigation is under the rankings table. Each week is preserved at a stable URL so you can reference specific weeks and compare trends over time.
How often is the ranking updated?
Once a week — every Sunday at 21:00 MSK a new edition is published. The page stays stable mid-week so it remains a reliable reference.