This week in AI
The AI news that matters for decision-makers — pulled from my aidevtube digest (≈45 channels) and re-lensed for business. Real developments, not trend-talk. Updated weekly.
Full digest on aidevtube — ~45 channels →
Anthropic releases Claude Haiku 5.5 with significant cost and performance improvements
Anthropic has launched Claude Haiku 5.5, a compact model positioned as a cost-effective subagent, priced at $0.10 per million input tokens. It significantly outperforms its predecessor, Haiku 4.5, with benchmark scores jumping from 15.7% to 72.4% on OSWorld. This development is particularly relevant for parallel agent architectures, offering a low-cost solution for software development contexts and verification tasks. For decision-makers, this model presents a strategic opportunity to reduce costs in AI-driven projects without sacrificing performance.
Reflection AI unveils Beam, a model optimized for coding tasks
Reflection AI has introduced Beam, a model with 501 billion parameters, 23 billion of which are active. It is specifically trained for coding and agentic tasks, boasting speed three to four times greater than comparable models. Founder Misha Laskin emphasized the need for a robust pretraining base for reinforcement learning to compete with Chinese models. This launch underscores the competitive space in AI development, highlighting the need for strategic investments in specialized AI capabilities.
Mistral Large 4 debuts as a sovereign AI model in Europe
Mistral has released Mistral Large 4, internally nicknamed 'LeChonk', featuring a trillion parameters with 49 billion active. It is trained entirely in European data centers, offering a million-token context and a top-five ranking in vulnerability research. While not leading globally, it is strategically positioned as an auditable tool for defense, law, and finance sectors. This model represents a significant step in European data sovereignty, offering businesses a compliant AI option within the region.
OpenAI introduces invisible watermarks for ChatGPT and Codex in the EU
OpenAI has implemented invisible watermarks in the EU for texts generated by ChatGPT and Codex. These are not hidden characters but statistical modifications in word choice, which remain detectable even after copying. Official detectors are not publicly available and are reportedly error-prone. This move highlights the increasing regulatory focus on AI-generated content, urging businesses to consider compliance and authenticity in their AI strategies.
Explore the potential of AI in filmmaking with TopView Film Studio
TopView Film Studio demonstrates how AI can bridge the gap between clip generation and targeted film scene direction. Using SeeDeng 2.5, users can create 30-second clips from 50 reference images, enhancing realism with microexpressions. This video is valuable for decision-makers interested in AI's role in creative industries, showcasing how AI tools can enhance control over film production without replacing human creativity.
▶ Watch · AI Filmmaking Academy on YouTube →This week's full digest — Claude Haiku 5.5, Mistral Large 4 et Beam : les modèles open source dominent la semaine — read all ~45 channels on aidevtube →