AI’s new capability frontier: from video generation to autonomous agents and price wars.

Another profound shift in the AI landscape is the rapid maturation of agentic AI – systems that don’t just analyze or chat, but can autonomously carry out sequences of actions. This week’s news offered a glimpse of how far these AI "agents" have come. Google’s Gemini release was framed explicitly as an “agent-first” move, emphasizing long-horizon tool use and complex task automation over simple chatbot conversations ([1]). And China’s tech giant Alibaba announced that its latest model, Qwen 3.7-Max, kept an AI agent running continuously for a staggering 35 hours, executing over 1,000 sequential operations without a noticeable performance drop ([2]). This was enabled by Qwen’s record-breaking 1-million-token context window, which allows the model to "remember" and process enormous amounts of information (equivalent to hundreds of pages of text) at once ([3]). In practical terms, these advances suggest that AI will soon be able to manage lengthy, multi-step business processes – think of an AI agent that can research a topic, draft a report, incorporate feedback, and refine the final output over many hours or days, all autonomously.

Such autonomous capabilities are not just theoretical. In a recent internal trial dubbed 'Project Deal', Anthropic deployed 69 AI agents that successfully negotiated and closed 186 sales transactions across an internal marketplace of over 500 products ([4]) – without direct human micromanagement. This experiment hints at a future where swarms of AI agents could handle routine business negotiations, customer interactions, or supply-chain optimizations in parallel, dramatically increasing efficiency. New players are also entering the fray: for instance, Elon Musk’s startup xAI just released a coding assistant called "Grok Build" that can coordinate up to eight parallel AI developer agents with access to a 2-million-token shared context window ([5]). In software projects, this means multiple tasks – from writing code and unit tests to performing code reviews – can be handled concurrently by specialized AI workers collaborating at unprecedented scale and speed.

However, as AI agents take on more autonomy, businesses must confront new risks. While current models are far better at extended reasoning and self-correction than earlier generations, they are still imperfect. A stray AI "hallucination" (a confidently wrong action or assertion) can have serious consequences if an agent is entrusted with executing financial transactions or managing sensitive data ([6]). Industry experts note that without proper safeguards, giving an AI agent free rein could turn a minor mistake into a major business liability ([7]). To build trust in these autonomous systems, companies are beginning to implement strong oversight mechanisms – for example, requiring human-in-the-loop controls where an agent’s plan must be reviewed and approved by a person before execution, and setting strict limits on what automated agents are allowed to do in enterprise environments ([8]). In short, the era of truly autonomous enterprise AI is on the horizon, but pairing innovation with governance will be essential to realize its benefits safely.

The past week also underscored how competition in AI is driving a rapid decline in the cost of advanced capabilities – and how open-source alternatives are expanding. Perhaps the clearest example is DeepSeek’s announcement on May 22 that its promotional 75% discount on the flagship V4-Pro model will be made permanent ([1]). This change slashes the price to just $0.435 per million input tokens (with $0.87 per million for outputs) ([2]), roughly one-quarter of the model’s original cost. At these rates, DeepSeek’s top model is now about 8× cheaper to run for generating input text and 10× cheaper for outputs than Anthropic’s Claude 4.7 model ([3]) – all while delivering competitive performance as one of the strongest open-weight AI systems available. Such dramatic cost reductions at the cutting edge mean enterprises can scale up AI deployments more affordably, analyzing larger datasets or running more extensive simulations without breaking the budget.

Meanwhile, the open-source AI community and smaller AI firms continue to chip away at the dominance of closed models. In late April, France-based Mistral released a 128-billion-parameter model (Mistral Medium 3.5) under a permissive open license ([4]), making one of the world’s largest language models freely available for anyone to adapt and deploy. This milestone – essentially delivering an open model at a scale that was once the province of only the largest tech companies – has fed early May’s discourse on how far community-driven AI can go. Meta, which set a precedent by open-sourcing Llama 2 in 2023, is expected to continue this strategy with an upcoming “Llama 5” model, although that next-generation model remains unreleased as of now ([5]). Even Alibaba has signaled it will release a smaller "Qwen Plus" model with open weights, complementing its closed Qwen 3.7-Max – a nod to the global demand for transparent and customizable AI tools ([6]). For enterprise users, the flourishing open-source ecosystem means more choice: businesses can potentially use and fine-tune cutting-edge models on their own infrastructure for greater data control and lower variable costs, rather than depending solely on the offerings (and pricing) of a few tech giants.

Still, industry leaders know that the most advanced capabilities often debut in proprietary models. Closed-source frontier systems from companies like OpenAI, Google, and Anthropic currently retain an edge in raw performance and unique features – for example, Google’s new Gemini Omni video-generation or OpenAI’s finely tuned instruction following. The strategic balance for enterprises will be to harness the best of both worlds. Many organizations may adopt a hybrid AI strategy: leveraging open-source models for use cases where they provide “good enough” performance or where data privacy demands self-hosting, and turning to proprietary cloud AI services when they need the absolute state-of-the-art power or specialized functionalities. This multifaceted approach can optimize costs and control while still allowing access to the latest AI breakthroughs.

If anything, the pace of AI advancement is set to increase, pushing the capability frontier even further. OpenAI’s forthcoming model (unofficially code-named “Spud”) is anticipated to mark a major leap, with company leadership hinting it will not be merely incremental ([1]) – implying a true “GPT-6” level advancement could be imminent. Google has similarly aggressive plans; a more powerful Gemini 3.5 Pro model is already in testing and slated for release in June ([2]), and Google DeepMind will likely continue rapidly iterating its Gemini series, broadening multimodal and reasoning abilities. Anthropic’s next upgrade (rumored as “Claude Mythos”) is expected to be its most powerful model yet ([3]), and we may see new entrants as well as regional players (such as Alibaba’s full release of Qwen) joining the top tier of AI within the year. In short, the race among AI model providers – both closed and open – will keep redefining what AI can do on ever shorter timelines.

We will also see AI systems entering new domains. The frontier is no longer confined to text or images on a screen; AI is making strides into the physical world. For instance, researchers recently demonstrated a robot control AI (Physical Intelligence’s π0.7) that can learn to combine skills and tackle entirely new tasks it wasn’t trained for ([4]) – a key step toward versatile, adaptable robotics. Such progress foreshadows the extension of AI-driven automation into manufacturing, logistics, and beyond. Likewise, with models like Gemini Omni starting to handle video, and others expanding into speech and sensory data, enterprises should prepare for AI that can design products, generate multimedia content, operate software systems, and perhaps even manage robotic processes, all using natural language commands.

Finally, companies must ready themselves organizationally for this coming wave. Adopting frontier AI is not just a tech upgrade but a strategic transformation. Many leading firms have already appointed Chief AI Officers to steer these efforts – 76% of organizations have a CAIO in 2026, up from just 26% a year before ([5]) – reflecting AI’s central role in business strategy. Moreover, 64% of CEOs now say they feel comfortable making strategic decisions based on AI-generated analysis ([6]), a dramatic increase in trust that underscores how pervasive and critical AI insights have become. In the next 6–18 months, the winners will be companies that proactively experiment with these new AI capabilities, invest in talent and infrastructure (and policies) to integrate them responsibly, and remain agile in adjusting strategy as the technology evolves. In an environment where your competitors, suppliers, and customers are all empowered by ever-smarter AI, keeping a pulse on the capability frontier isn’t just a tech task – it’s a leadership imperative.