
Tech • IA • Crypto
OpenAI unveiled GPT-5.6 in a restricted preview, limiting access to select partners with no public release timeline. The model introduces a three-tier system—Sol, Terra, Luna—with Sol Ultra reportedly using sub-agent architectures for complex reasoning. OpenAI claims up to 3x token efficiency gains while maintaining top-tier performance. The cautious rollout reflects growing alignment with governments and heightened concern over deploying powerful systems broadly.
Anthropic launched Claude Design 2.0, redesigning its interface around a prompt-first, template-driven workflow. A major addition is direct canvas editing, allowing users to modify designs without repeated prompting, reducing token usage. Usage limits are now unified across Claude Pro and Max plans, simplifying access across tools. The platform also introduces design-to-code sync with Claude Code, tightening integration between design and development.
Z.AI’s GLM 5.2 is rapidly gaining traction as an open-weight model competing with leading U.S. systems. It ranks among the top 10 most-used models and shows strong performance in cybersecurity tasks like vulnerability detection. Its open distribution increases accessibility but raises concerns about misuse and lack of oversight. The release highlights accelerating competition between Chinese and U.S. AI ecosystems.
Microsoft CEO Satya Nadella is advancing the concept of “token capital,” framing AI usage as a core economic resource. This includes compute access, proprietary data, and internal learning systems that accumulate advantage over time. As models commoditize, differentiation shifts to learning loops and retained knowledge. Firms that effectively compound token usage into proprietary systems gain durable competitive leverage.
Major firms including Amazon, Walmart, Uber, and Meta are now restricting internal AI usage due to rising costs. Token consumption has become a primary expense driver, turning AI into a managed operational budget similar to cloud infrastructure. What began as experimentation is transitioning into disciplined allocation. This shift signals that scaling AI usage economically remains an unresolved challenge.
Traditional benchmarks are losing credibility as primary measures of model quality. Platforms like LM Arena emphasize user preference testing, offering more practical insights into real-world usefulness. Models such as GLM 5.2 and Claude 3.5 variants show differing results between benchmarks and user rankings. The shift reflects growing demand for evaluation methods that mirror actual usage პირობ.
AI-generated search summaries are set to roll out in France, expanding a feature already deployed in other regions. These overviews prioritize synthesized answers over traditional link-based results, reshaping user interaction with search engines. The change could significantly impact traffic distribution across websites. It also reinforces the trend of AI systems becoming primary information intermediaries.