Foundation models push the capability frontier forward.

Multimodal AI – the ability for models to handle and generate multiple forms of media – made a significant leap with Google’s Gemini Omni. This new system can create high-fidelity video from virtually any input; for example, users can feed it a combination of text descriptions, images, or audio and have the AI generate a realistic custom video ([1]). Early uses are creative (e.g. producing marketing videos or training materials on the fly), but the technology has clear enterprise potential. Companies could soon leverage such tools to automate video content creation, simulate real-world scenarios for training or design, and communicate ideas in richer media formats – all without large specialist teams.

AI is also becoming more agentic, meaning these systems increasingly take autonomous actions on behalf of users. In a striking Google I/O demo, the new Antigravity coding agents (powered by Gemini 3.5 Flash) collaborated to build a simple operating system in just 12 hours ([2]). The automated dev “team” consisted of 93 AI subagents working in parallel on 15,000 software tasks and processing 2.6 billion tokens – all for under $1,000 in cloud compute costs ([3]). While this was an early experimental showcase, it foreshadows how AI could tackle complex projects at a speed and cost efficiency impossible for human teams, potentially revolutionizing software development and other labor-intensive knowledge work.

Established players are investing in research to boost these autonomous capabilities as well. Microsoft researchers this week open-sourced a training method called ECHO that helps AI agents learn from intermediate feedback (like reading error messages from a computer terminal) instead of only from end results ([4]). In tests, ECHO nearly doubled agents’ success rates on coding tasks and achieved the same performance with 2.3× less training time ([5]). Advances like this will likely yield more capable and reliable AI agents that businesses can deploy to automate multi-step processes (such as IT operations or data analysis) with greater confidence and speed.

Even AI’s top architects are envisioning more radical forms of autonomy. OpenAI’s CTO Mira Murati, for example, has discussed AI that can 'listen, watch, interrupt, use tools, and keep collaborating in real time' ([6]) rather than waiting passively for human prompts. In other words, future AI assistants could operate continuously and proactively in the background, functioning as collaborative digital colleagues rather than just reactive tools. If this vision comes to fruition, it will require enterprises to rethink how work is organized and supervised, as employees begin to delegate more complex, ongoing tasks to always-on AI co-workers.

Staying at the AI capability frontier is not just a technical race – it’s also a financial and strategic one. This week, Anthropic (the creator of the Claude model) announced an eye-popping funding round of about $30 billion that would value the company at over $900 billion ([1]) – surpassing OpenAI’s valuation from March 2026. According to Anthropic’s CEO, most of this capital will go toward massive cloud compute contracts with providers like AWS and Google Cloud ([2]) – reinforcing the maxim that 'whoever controls compute controls model capability' ([3]). In practice, dominating the next wave of AI will require unprecedented access to computing power. Enterprises should take note: partnering with well-resourced AI vendors (or investing in dedicated AI infrastructure) is becoming crucial to leverage the most advanced models.

The cost of computation is already a rising factor. Skyrocketing demand for AI has begun straining hardware supply and budgets – even year-old NVIDIA H200 chips are now renting for more than newer models due to GPU shortages and surging usage ([4]). Tech giants are responding with astronomical spending; Google, for example, expects to invest $180–$190 billion in AI development this year alone ([5]). In the near term, businesses should be prepared for higher prices and potential delays in accessing cutting-edge AI capabilities as competition for limited compute intensifies. Ensuring reliable access to AI hardware (through cloud providers or on-premises investments) is becoming a key aspect of competitive strategy.

Meanwhile, open-source AI efforts are rapidly closing the gap – and driving down costs. Researchers in China have introduced open models like DeepSeek V4 (a 128 billion-parameter model with a one-million-token context window) that already achieve parity with the previous generation of leading U.S. systems on reasoning benchmarks ([6]). Meta’s attempt to deliver a similar open-release model, code-named Avocado, has been delayed to next month after internal tests showed it still fell short of top performers like GPT-5.5 and Claude ([7]). But other players are pressing forward: Google DeepMind recently released Gemma 4 – its most advanced open-weight model family yet – under an Apache 2.0 license ([8]), focusing on high 'intelligence per parameter' efficiency for both cloud and edge deployments ([9]). And Elon Musk’s xAI has likewise streamlined its roadmap, retiring eight legacy “Grok” models to concentrate on its new agent-centric flagship, Grok 4.3 ([10]). The takeaway: enterprises will soon enjoy a greater choice of high-end AI models beyond the big proprietary platforms, offering more flexibility to customize and deploy AI on their own terms.

Perhaps most striking, many of these open models undercut the expenses of proprietary AI by orders of magnitude. DeepSeek V4’s usage price is around $1.74 per million input tokens, compared to $30 for OpenAI’s GPT-5.5 ([11]) – roughly a 98% cost reduction. And yet the performance gap between DeepSeek and the top closed models on broad question-answering tests is only a few percentage points ([12]). For businesses, this is a potential game-changer: open-source AI can enable massive cost savings and deployment control. It may also pressure commercial AI vendors to lower prices or increase value, as customers gain viable low-cost alternatives that deliver near-frontier capabilities.

One immediate effect of the advancing capability frontier is a push to broaden AI adoption across organizations of all sizes. Despite contributing nearly half of U.S. economic output, small and mid-sized businesses have been slow to embrace advanced AI – only about 7% of U.S. SMEs have adopted cutting-edge AI tools so far ([1]). To change that, Anthropic this week launched Claude for Small Business, which packs 15 ready-made AI workflows into popular software used by smaller firms (from QuickBooks to Google Workspace) ([2]). By embedding AI assistants into everyday tasks like bookkeeping, invoice processing, marketing content creation and more, vendors aim to dramatically increase AI uptake in this huge segment. Larger enterprises should also pay attention: as advanced AI becomes accessible to even the smallest players, disruptive competition could increasingly emerge from unexpected quarters.

At the same time, leading corporations are rapidly scaling up their internal use of AI. Global consultancy PwC, for instance, is deploying Anthropic’s Claude model across its 300,000-strong workforce as part of a new strategic alliance ([3]). In pilot projects, PwC slashed an insurance underwriting process from 10 weeks to 10 days using Claude, and saw other processes run up to 70% faster from end to end ([4]). These tangible results show that today’s generation of AI tools can dramatically accelerate knowledge work and operations. Enterprises that move early to implement and learn from such AI deployments – including establishing internal AI centers of excellence to train employees – stand to gain a significant efficiency and innovation edge.

As AI systems gain the ability to plug into more sensitive business data and workflows, companies must update their governance strategies. For example, OpenAI’s new finance plugin now allows ChatGPT to connect directly with bank and investment accounts via the Plaid API ([5]), letting the AI analyze financial transactions and balances for real-time insights. This unlocks powerful use cases (like automated cash-flow analysis or expense management), but also raises critical questions around data security and regulatory compliance ([6]). Business leaders need clear policies and robust safeguards for AI handling of confidential information – balancing the drive to innovate with the need to protect privacy and comply with laws.

Finally, the breakneck pace of AI advancement means that capabilities considered cutting-edge today will be commonplace within 6–18 months. Many of this week’s breakthroughs – from multimodal content generation to autonomous agents with long-term memory – are likely to become standard features of enterprise software in that timeframe. To stay ahead, organizations should be mapping these technologies to their strategic plans now. That includes identifying high-impact areas for AI adoption, investing in the necessary data infrastructure and skills, and engaging with both major AI providers and open-source solutions. By proactively embracing the new AI frontier, companies can drive innovation and efficiency in their operations – and avoid being left behind by faster-moving competitors.