Siri AI Open to Developers, OpenAI Claims 250 Million Users

The week combined product announcements and scaling indicators in AI. OpenAI reports a massive user base while developer tools are multiplying. Work on provable inference and hardware plans outlines distinct technical directions.

OpenAI claims 20% of Google search and 250 million users

OpenAI states that searches via ChatGPT account for about 20% of those on Google, and reports 250 million people interacting with its service. An image generation model from text has been trained for $21,000, providing a benchmark for such projects. The actual costs of tokens, a central issue for sizing usage and billing, are also detailed in a dedicated analysis.

New developer tools: Siri AI, Gemini Notebook, Baseten, Fireworks

Siri AI is now accessible to all developers. Meanwhile, the Gemini Notebook API has been introduced, featuring new capabilities to integrate AI tools into applications. On the platform side, Baseten presents Grid, Sandboxes, and Serverless Loops, while Fireworks launches faster inference tools and offerings for enterprises.

Research, methodology, and technical roadmaps

Advancements have been observed in the field of provable inference applied to large language models, which is likely to enhance the transparency and verifiability of decisions generated by AI systems. The impact of reasoning efforts on the resulting score and associated costs has been analyzed, and lessons have been learned from auditing public evaluations. It is considered misleading to claim that Astra can read and write weights to optimize Qwen. DeepSeek plans to train its next model on Huawei Ascend chips, and ChatGPT is set to start creating GIFs. Insights on balancing AI value and risk come from the Gartner IT Symposium/Xpo 2026, and the acceleration of engineering for the physical world is highlighted. Finally, the need for memory for agents and the Agent State Machines approach, which proposes a deterministic structure for agents, are also discussed.