Coding models and shaky benchmarks & Enterprise AI moves on deployment - AI News (Jul 10, 2026)

[0:32] Coding models and shaky benchmarks

We'll start with the coding race, where the headline is not just new models, but whether the scorecards themselves can be trusted. Cognition has launched SWE-1.7, its strongest software-engineering model so far, and the notable part is not only capability. The company says it got there with a more efficient reinforcement-learning pipeline and better support for long, real-world coding sessions that can stretch for hours. That matters because it pushes back on the idea that reinforcement learning has already tapped out, and it suggests coding agents may keep improving without infrastructure costs exploding.

In AI interfaces, OpenAI is rolling out GPT-Live, a new voice system for ChatGPT that is designed to feel much more like a real conversation. The big change is that it can listen and speak at the same time, rather than waiting for rigid turns. It can also hand harder tasks off to a more powerful model in the background while keeping the conversation flowing. If that works as promised, it moves AI assistants closer to something people might actually talk to throughout the day, instead of only using them as a text box.

A very different example comes from Ello, which explained how it built a real-time AI tutor for children aged four to nine. The important point here is that model intelligence was only part of the challenge. The company says it had to redesign the whole system around speed, safety, and teaching strategy, because kids will not wait around for an agent to think. That is a useful reminder for the whole industry: in sensitive settings like education, product design and behavioral control can matter just as much as the model underneath.

In robotics and research, Mistral has introduced Robostral Navigate, an 8 billion parameter model that lets robots follow natural-language instructions using just a single RGB camera. It reportedly beats earlier single-camera systems and even some setups with more sensors. If those results hold up, that could lower the cost of useful robot navigation and make deployment more practical in warehouses, hotels, factories, and other everyday environments.

And in one of today's more unusual research stories, scientists at EPFL unveiled NEvo, a system that generates videos specifically designed to activate chosen regions of the visual brain. In effect, AI is being used to create custom visual stimuli for neuroscience. The work reportedly outperforms standard handcrafted clips and shows just how important motion can be in triggering higher-level visual responses. It matters because this is AI not just making media, but helping researchers understand how perception itself is organized.

Finally, on safety and long-term direction, Anthropic and AE Studio introduced GRAM, a research method for isolating sensitive capabilities inside removable modules. The idea is to separate knowledge from areas like cybersecurity or virology so it can be disabled or removed later without retraining an entire model. It is still early research, but it points to a more flexible approach to model governance than an all-or-nothing release strategy.

That connects with a broader thread in the field. A new taxonomy of self-evolving agents maps out systems that can build reusable skills or even search for better ways to improve themselves over time. And a separate proposal called Plan A argues that the world should slow the push to superintelligence, spread AI capabilities more widely, and avoid letting too much power collect in the hands of a few actors. You do not have to agree with every detail to see the underlying issue: AI is no longer just a product story. It is becoming a question of who gets to shape the next operating system for work, security, and society.