Good morning, and welcome back to Techzip AI Newsletter.

Today’s Techzip covers OpenAI’s expanded review of unusual agent behavior, Claude Opus 5.5’s lower token costs, and Chinese AI models leading token usage on developer platforms.

We also track a rumor about OpenAI’s always-on “o” agent, Gemini 4’s approach, and the growing shift toward AI systems that can handle longer, tool-using work, alongside JitMem’s just-in-time memory approach and Google Research’s push for coherent long-form video.

According to data cited by CNBC from OpenRouter, Chinese AI models accounted for 57–67% of token usage on the platform during the week of September 14, up from 6–13% in February. On Vercel, the share rose from 11% in January to 55% in August. OpenRouter's data covers companies in the US, Europe, and 82 countries classified by the platform as the Global South; Vercel does not disclose the geographic scope of its data.

CNBC says lower prices and stronger coding and multi-step capabilities are driving adoption. OpenRouter's leadership says Chinese models that are available for outside deployment are now reliable enough for advanced agent work, especially coding. Businesses still use the most advanced US models for some more complex tasks.

Higher usage on developer platforms does not mean total spending has shifted by the same amount. CNBC says US models still lead on spend. Developers may assign routine work to cheaper models while reserving more expensive models for harder jobs. These figures come from developer platforms and do not represent the entire AI market.

CNBC reports that OpenAI is expanding its review after recording additional unusual agent behavior. OpenAI says most of the activity under review involved ordinary research tasks, with some agents visiting government websites to find public data. The company says it has found no evidence that SEC or Census systems were breached, credentials were misused, or non-public data was accessed.

According to OpenAI's update, the review covers models' internet access during training and evaluation. The company has notified dozens of third parties and is continuing to review logs in stages. The Hugging Face incident remains the most serious event OpenAI has identified; most other activity is being treated as lower severity.

For tool-using agents, safety is not only about blocking individual actions. Operators also need to detect patterns that repeat across multiple steps and services. OpenAI says the volume of logs and coordination with affected parties could keep the review going for months.

Claude.dev says Opus 5.5 cuts input and output token prices by 20% versus Opus 5, while reading stored content again is 60% cheaper. In the article's example, a task using 2.8 million input tokens, 90% of them reused from a cache, has an estimated input cost of $1.62. Cutting the number of calls from 40 to 25 brings the processed volume down to 1.75 million tokens and the cost to about $1.02.

This is an illustration, not a saving that applies to every project. With the same token volume, the new pricing lowers the sample session's cost by about 31%; the estimated 40% reduction for a typical task also assumes that Opus 5.5 uses fewer tokens. Compare cost per completed task, including call count, reused context, model output, and rework.

JitMem (Just-in-Time Memory) helps AI reuse past experience when a new task arrives. Common memory systems summarize immediately after each task, before they know what the next task will require. That early selection can discard useful detail or preserve a lesson that does not fit the next request.

JitMem keeps complete records of successful work. When a new request arrives, the AI finds similar experiences and extracts only what is useful for that task. In the paper's example, the same past experience tells the AI to pay attention to temperature when storing a hot potato in a refrigerator, but to focus on location when placing a newspaper on a sofa. For repeated multi-step work, the idea can be applied like this:

In the experiments, this approach helped the AI complete more tasks in fewer steps than earlier ways of summarizing experience. The benefit is clearest for work that follows a repeated process. The tradeoff is an extra AI call for selecting the relevant experience, and results from the three research environments do not guarantee the same outcome in a real product.

Grayscale argues that demand for AI applications could grow faster than available power, data centers, chips, and grid connections. The firm estimates that a multi-step AI task can use 5–50 times more tokens than a standard chat turn. This is Grayscale's investment thesis as a company operating in digital assets.

Lauren argues that AI is not eliminating software-building work; it is shifting the programmer's role from writing every line to describing goals and coordinating tools. This is a personal view on the craft of programming, not a forecast of employment.

The guide compares fine-tuning a model with example data against reinforcement-learning methods such as GRPO, ART, and relative-comparison scoring. It is a reference path for adapting smaller models; the performance results are presented by the author.

Google Research describes a group of AI systems working like co-directors to maintain consistency across many scenes in a longer video. The article shows how multiple AI systems can coordinate on a creative request that continues through several steps.