🚀 Modèles & Releases
Anthropic has launched Claude Sonnet 5 (claude-sonnet-5) with introductory pricing of $2/$10 per MTok through August 31, 2026, supporting a 1M token context window and 128k max output tokens. (Source)
AionLabs released Aion-3.0-Mini, a multi-model roleplaying and storytelling system built on the DeepSeek family of models. (Source)
Tencent released the Hy3 model, a 295B-parameter Mixture-of-Experts model for reasoning and agentic workflows. (Source)
🔬 Recherche
FreqDepthKV, a new inference-time KV cache compression method, was introduced to address memory and bandwidth costs in long-context LLM inference while preserving layer-specific evidence. (Source) MANCE (Manifold Aware Concept Erasure) is a new method proposed for removing a target concept from a representation while preserving other correlated information. (Source) A paper on Object-Evidence Preserving Token Merging for vision-language retrieval aims to reduce storage and scoring costs without removing crucial object-level evidence. (Source)
🏢 Industrie & Business
Meta launched a new AI image generator, Muse Image, which has already faced user pushback over the use of their photos. (Source) Microsoft is joining the AI cost-cutting trend by relying more on its own models. (Source) A TechCrunch analysis suggests the rise of open-source AI isn't yet hurting Anthropic, as both capture different phases of the same life cycle. (Source)
🛠️ Outils & Produits
Google DeepMind announced an expansion of Managed Agents in the Gemini API, including features for background tasks and remote MCP. (Source) Ollama released version v0.31.2, which includes fixes for CUDA toolkit lookup and parallel processing. (Source) OpenHands released version 1.9.0, featuring semantic file chunking using tree-sitter AST parsing and an organization conversation admin dashboard. (Source)
⚙️ API & Dépréciations
Anthropic's Claude Sonnet 5 is available at a discounted introductory rate until August 31, 2026, after which standard pricing of $3/$15 per MTok will apply. (Source)
⚖️ Régulation & Société
Illinois Governor JB Pritzker signed what is described as the country's strongest AI regulation bill, aiming to mitigate risks. (Source) The EU AI Act received its final green light from the Council. (Source)
💡 À retenir
Anthropic's Claude Sonnet 5 launch with discounted pricing provides a clear, time-limited opportunity for cost optimization before rates increase. The new KV cache compression research (FreqDepthKV) is critical for production teams scaling long-context applications, as it directly tackles a major inference bottleneck.