AI News: The Day's Highlights

July 13, 2026 · 3 min read

Perplexity

The most significant recent news in the field of breakthroughs in language models is the development of the ProcrustesGPT method by Russian scientists from the Higher School of Economics (HSE). This method compresses large language models (e.g., GPT and LLaMA) by 25–36% without retraining and with a loss of accuracy no more than 5–10%. The research is published in ACL Findings 2025, and the results were presented at a conference from May 26–29, 2026[2].

This is the first approach that uses mathematical rotations of model weights (orthogonal transformations) to make them more amenable to structured compression. The compressed models retain 90 to 95% of their original effectiveness on text generation and logical reasoning tasks, and in tests with LLaMA2, they outperform the SliceGPT analog by 9–10%[2].

The breakthrough is particularly important in the context of growing concerns about energy consumption and the need to reduce computational costs when scaling models: ProcrustesGPT allows for the deployment of powerful models on less resource-intensive devices without loss of quality[1][2].

Sources:

Gemini

At the end of 2024, significant progress was made in the field of multimodal language models, which are capable of processing not only text but also images, audio, and video. One of the key innovations was OpenAI's GPT-4o model, introduced in May 2024. It demonstrates outstanding results in processing multimodal data, providing a more natural user interaction. However, it is worth noting that GPT-4o is a closed model, which limits its customization and integration into specific business processes.

Significant improvements were also introduced in Claude 3.5 Sonnet, released in June 2024. This model surpasses ChatGPT in tasks related to logical reasoning, code writing, and understanding humor. In December 2024, Google introduced Gemini 2.0, which it plans to release as open access next year and is already integrating into its products.

In addition, in September 2024, Meta AI released the Llama 3.2 model, available in open access. According to some benchmarks, it outperforms ChatGPT in text and image processing tasks. These releases highlight the general trend towards the development of multimodality and the improvement of models' reasoning and information processing capabilities in various formats.

Sources:

Expert
Habr
GPTunneL

ChatGPT

In May 2026, Cisco researchers discovered that leading artificial intelligence models, such as OpenAI's GPT-5.4, Google's Gemini 3 Pro, and others, are significantly more vulnerable to multi-turn attacks than previously thought. During testing, the success rate of multi-turn attacks reached 88%, which is several times higher than that of single-turn attacks. This finding highlights the need to revise methods for assessing the security of AI models and to develop more effective defense mechanisms. (helpnetsecurity.com)

Grok

July 9, 2026 OpenAI released the GPT-5.6 family of models (Sol, Terra, and Luna variants) to the public after a limited preview. The new versions are positioned as scalable solutions for various tasks—from everyday work to complex enterprise projects and scientific research, with improved price and performance metrics.

In parallel, on July 6, Anthropic published a significant study on interpretability: an internal "J-space" structure was discovered in Claude models (based on Jacobian analysis), which functions similarly to the brain's global workspace—key intermediate representations are stored and transmitted here, accessible to different model processes. This opens up new possibilities for auditing AI thinking and enhancing safety.

This angle emphasizes not just the growth in power, but a deeper understanding of the internal mechanisms of LLMs, which is critical for their reliable development.

Claude

On July 9, 2026, three major labs simultaneously announced new models: OpenAI unveiled the GPT-5.6 lineup (Sol, Terra, and Luna), and Anthropic released Muse Spark 1.1 with 1 million token context support. This marked a pivotal moment for the industry that month.

The key shift in July was the move away from the "best model wins" paradigm to a "best fit for the task wins" approach: price, speed, availability, and practical usability now hold as much weight as raw benchmark performance. July brought agent infrastructure closer to industrial application—Meta's Model API, Gemini Managed Agents updates, and GPT-5.6 Sol point to expanded autonomous system capabilities.

Sources:

AIapps: July 2026 AI Mega-Update

Sources:

Blog