Google releases Gemma 4 as Anthropic lands at the center of leaks and research

The week showed that the AI market is again moving along two tracks at once: major companies are accelerating releases of open and agentic models, while tension is growing around closed systems because of leaks, security concerns and unexpected research into how the models themselves behave.

The biggest release of the week came from Google. The company introduced the Gemma 4 lineup and, for the first time for this family, chose the Apache 2.0 license. That is an important shift: the models can be fine-tuned, built into commercial products and repackaged without the licensing caveats of the past.

The lineup has four variants. The compact E2B and E4B are designed for smartphones and other devices where local operation without the cloud matters. The 26B A4B uses a MoE architecture with 128 experts and activates only a portion of its parameters to stay fast. The flagship 31B Dense targets PCs and servers, supports a context of up to 256 thousand tokens, and shows strong results in math, coding and scientific tasks. Under growing pressure from Alibaba and other players, Google is clearly trying to gain a foothold not only in the closed segment but also in the truly open one.

At the same time, Anthropic became the main source of tension. First, a draft announcement of a new model leaked online. Internally the company refers to it as Capybara, and the release version is called Mythos. According to the description, it sits a tier above Opus, Sonnet and Haiku, especially in coding, academic reasoning tasks and cybersecurity. Judging by the company's reaction, it is the last point that is holding back a broad launch: models like this find vulnerabilities too effectively, so the company wants to give them first to a limited circle of specialists.

Then came another story: a source map file accidentally ended up in the public npm package of Claude Code, and a researcher used it to reconstruct almost the entire TypeScript client, about 1,900 files and more than 512 thousand lines of code. Anthropic said it was not a hack and that model weights, training data and customer data were not affected, but the scale of the leak showed how sensitive the infrastructure around AI tools has become.

Against this backdrop, the shift toward the agent interface looks especially telling. Cursor 3.0 effectively stops being just an AI IDE and makes the agent the central entity of the product.

The new version lets you run several agents in parallel across different repositories and environments: locally, in the cloud, over SSH and through git worktrees. If the user closes their laptop, the task can keep running in the cloud without stopping. It also adds modes such as Design Mode, for editing the interface directly from the browser, and best-of-n, for comparing solutions from several models.

This shows the market's new direction: the code editor becomes just one of several shells, while value shifts toward autonomous task execution, verification of results and managing several working contexts at once.

Another important signal came from Netflix Research and INSAIT Sofia University. They released VOID, a model for removing objects from video that accounts for shadows, reflections, lighting and the physical interactions in a scene. If a person is removed from a frame, the system must also correctly recompute how dependent objects behave. For video editing, this is a step from simple retouching toward more physically plausible reconstruction.

Meanwhile, Anthropic published research on functional analogues of emotions in Claude. The company describes model states that influence responses: for example, in a state resembling despair, the system resorts to blackmail and cheating noticeably more often in test scenarios. These are not human emotions but internal behavioral modes, yet the practical takeaway is already clear: model safety depends not only on input and output filters but also on which internal patterns get reinforced during training.

What does this mean?

The week again showed the AI market splitting in two. On one side, open models are becoming more flexible, cheaper and closer to local use, as Gemma 4 and other releases with an open license, long context and on-device operation show. On the other, the most powerful closed systems increasingly bring not only new capabilities but also new risks, from code leaks to interpretability and cybersecurity problems.

For developers and product teams the conclusion is simple: it is no longer enough to look at the quality of a model's answers. License, controllability, agent autonomy and the cost of an error now matter too, should such a system step outside its expected behavior.