August 5, 2026

Meta released Muse Spark 1.2 and the agent Muse Code at 20 cents per million tokens, acknowledges shortcomings in its own benchmarks

Meta released the coding model Muse Spark 1.2 and the agent Muse Code (beta), with a plan starting at 20 cents per million output tokens in exchange for data sharing. The company itself acknowledges that its comparative benchmarks against competitors are not methodologically balanced.

New: Muse Spark 1.2 pricing: 20 cents per million output tokens; The model compresses context instead of truncating it in long sessions; Greater compute investment in training on programming tasks; Meta acknowledges methodological shortcomings in its own benchmark tests; Muse Code can restore context after a failure in long-running tasks

DeepMind changes leadership: Demis Hassabis steps away from operations, replaced by Koray Kavukcuoglu, following the departure of the founders of the startup Discovery Loop

Demis Hassabis stepped down from operational leadership at DeepMind, and Koray Kavukcuoglu became the new leader; researchers from reinforcement learning teams continue to leave the company. This follows the departure of Jeff Dean, Sanjay Ghemawat, Oriol Vinyals and Quoc Le, who founded the startup Discovery Loop.

New: Demis Hassabis is stepping away from operations at DeepMind; Koray Kavukcuoglu is becoming the new leader of DeepMind and the remaining co-leader of Gemini; More senior talent in reinforcement learning is leaving DeepMind

Mobileye deployed an AI agent on Amazon Bedrock AgentCore to automate internal support

Mobileye deployed an AI agent for internal support on Amazon Bedrock AgentCore that, according to the company, reduced response time by 90 % and achieved accuracy above 95 %, eliminating 66 % of routine manual queries about the processing status of driving data.

Overloaded US power grid: AI is becoming a necessity for management, IEEE launches training program

The US power grid is overloaded by growing demand from data centers (220 GW of new connection requests in Texas) and fluctuating generation from renewable sources. According to experts, AI has become a necessity for grid management; IEEE has launched a course to train engineers in this field.

Mistral has released Shieldstral, an open model with 3 billion parameters for checking content safety

Mistral has released the open-weight model Shieldstral (3 billion parameters) for classifying the safety of text and images. According to the company, it achieves an F1 score of 84.9 % on text (on par with the seven-times-larger GPT-OSS-Safeguard-20B) and 83.8 % on images. Rules can be configured at runtime in natural language without training.

Reddit introduces Rules Hub, community moderation using language models

Reddit is launching Rules Hub, an LLM-based moderation tool intended to replace the existing keyword matching in Automoderator. It was tested for several months across 700+ communities, and the company plans to make it fully available later in 2026.

US government creates a classified framework for security reviews of frontier AI models

According to Tech Policy Press, the US government is developing a classified framework under Executive Order 14409 for reviewing frontier AI models before release, without published criteria. The process led to a 19-day block on models from Anthropic and a limited launch of GPT-5.6 from OpenAI.

US government introduces voluntary framework for testing risky closed-source AI models, excludes open models

According to Axios, the Trump administration has introduced a voluntary framework for testing the security risks of AI models. It applies only to closed-source models with cutting-edge capabilities that pose a national security risk; open models are excluded, and key terms are undefined.