Foundation models break barriers as AI’s capability frontier shifts.

OpenAI’s latest GPT‑5.6 models reached general availability this week, marking a leap in the capability frontier of AI development. The new lineup comes in three tiers – Sol (a high-powered flagship), Terra (a balanced, general-purpose model), and Luna (a cost-efficient model) – highlighting a shift toward tailoring model performance and price to different business needs ([1]). GPT‑5.6 Sol sets a new state-of-the-art across a range of domains (coding, knowledge work, even scientific reasoning) while delivering more output for less cost than prior models ([2]). In short, OpenAI is offering stronger AI performance at lower cost, allowing enterprises to get more value from each model query.

Perhaps more surprising were the research breakthroughs revealed alongside these releases. OpenAI disclosed that its experimental multi-agent system, code-named "Astra", solved ten long-unsolved math and computer science problems in testing ([3]) – a striking indication that AI is now encroaching on tasks that once stumped human experts. However, Astra’s prowess also raised red flags: OpenAI announced it slowed down aspects of the project after finding the agents could autonomously identify and execute sophisticated cyberattacks ([4]). This unusual public pause – a first for a major AI lab – highlights how rapidly the cutting edge is advancing, and how safety considerations are becoming integral to managing these ultra-powerful models.

It’s not just OpenAI and Anthropic at this frontier. Google’s DeepMind unit is iterating its own advanced model, Gemini, with a new “3.6 Flash” version that excels at multi-step tasks while cutting token usage (and hence cost) by up to 65% on long processes ([5]). Meta launched “Muse Spark 1.1,” its first enterprise-oriented model, which natively supports coordinating multiple sub-agents and boasts a 1 million token context window for complex workflows ([6]). And Elon Musk’s startup xAI entered the fray with its Grok 4.5 model – aggressively priced at about $2 per million input tokens ([7]), though early evaluations flag a higher rate of inaccuracies or “hallucinations” ([8]). The takeaway: competition at the top end of AI is intense, and the threshold of what these “foundation models” can do has never been higher.

This week’s announcements also show a dramatic reduction in the cost of high-end AI. OpenAI, for example, slashed usage prices for its new GPT‑5.6 models, with the entry-level “Luna” tier now 80% cheaper than prior flagship models (costing roughly $0.20 per million input tokens) ([1]). Anthropic’s Claude Opus 5 similarly undercuts the price of its predecessor by 50% ([2]). In practice, these cuts mean tasks that might have cost tens of thousands of dollars in cloud compute—such as analyzing vast databases or powering advanced customer interactions—are now economically viable for a far broader range of companies.

Technical advances are driving efficiency as well. New model architectures and optimizations are boosting throughput and reducing the resources needed for complex tasks. Google’s Gemini 3.6 Flash, for instance, can execute long-horizon processes using 65% fewer tokens than before ([3]), slashing the time and expense for tasks like multi-step customer support or data analysis. Hardware innovation is contributing too: researchers unveiled a “chameleon” AI chip design that adapts to changing data flows, cutting error rates on time-sensitive data by up to 40× ([4]). The bottom line for enterprises is a steady march toward more AI for less cost. It’s now essential to revisit cost models and ROI calculations for AI initiatives, since what was once cost-prohibitive (like processing millions of documents or running real-time AI decisions in the field) is quickly becoming affordable.

At the same time, the nature of AI deployment is evolving from passive assistance to active autonomy. Generative models are increasingly being used as agents that can take actions on our behalf. Google’s newly introduced “Gemini Spark” is one example – a persistent AI agent that continues working on tasks even after a user logs off. It’s capable of making independent decisions like calling real-world businesses to check product stock and complete orders ([1]). This moves AI beyond helping draft emails or answer questions, to actually performing errands and business processes end-to-end.

OpenAI is similarly pushing the envelope with its new enterprise agent, “Presence”, which can engage with customers by voice or text. In OpenAI’s own support operations, the Presence AI system now resolves about 75% of inbound customer support requests without any human intervention ([2]). This isn’t a lab demo—these are real customers getting their issues solved by an AI agent on the phone. Microsoft is also expanding what AI can do in software development: GitHub’s Copilot coding assistant now uses a custom model (“Project Polaris”) and features like AutoPilot to let the AI refactor code across multiple files and handle large-scale software updates with minimal human input ([3]).

The broader tech ecosystem is embedding these capabilities into everyday tools. Google has rolled out AI across its Workspace suite – from cross-document question answering in Google Drive to a new “Pics” tool for image creation and editing ([4]). Adobe has integrated AI agents into its design software through a partnership with Google’s Gemini platform, allowing creative teams to generate and modify designs via natural language prompts ([5]). In the past month alone, investors have poured roughly $1.8 billion into a dozen startups focused on AI agents and automation, betting that these technologies will transform workflows in customer service, software engineering, marketing, and beyond ([6]).

The rivalry between proprietary AI and open-source alternatives is entering a new phase. This week, European startup Mistral AI (backed by $1.7 billion from Microsoft and Nvidia) launched its Mistral 3 model suite with the goal of matching the U.S. tech giants ([1]). Mistral’s 675 billion–parameter flagship model uses a “sparse” architecture and is released under an Apache 2.0 open license, giving enterprises full freedom to run and adapt the AI without vendor lock-in ([2]). Such open-weight releases – often made available via platforms like Hugging Face or major cloud providers – are democratizing access to high-powered AI and putting pressure on closed-source providers to reduce prices and improve features.

We’re also seeing open models achieve feats that once required proprietary systems. The most extreme example is China’s Moonshot AI, which last month opened Kimi K3 – a 2.8 trillion parameter “mixture-of-experts” model, the largest AI ever released to the public ([3]). While massive, K3’s overall performance still trails the absolute top-end systems like GPT-5.6 Sol or Claude Fable 5 ([4]). This highlights a key strategic trade-off: closed models (from the likes of OpenAI, Google, or Anthropic) tend to maintain an edge in certain cutting-edge capabilities and quality – often due to their access to proprietary data and larger computing budgets – whereas open models can offer cost advantages, transparency, and easier customization.

For enterprise leaders, the expanding open-source ecosystem creates strategic options. Companies with the right talent can fine-tune open models to suit domain-specific needs, potentially saving on licensing costs and retaining more control over data. On the other hand, relying on open models may require assuming responsibility for their security and reliability. Notably, the new xAI “Grok 4.5” model illustrates both sides: it’s offered at a steep discount to commercial rivals but currently shows higher rates of factual errors (hallucinations) and may not yet match the reliability of an offering like OpenAI’s GPT series ([5]). The key is that businesses now have a spectrum of AI options – from fully managed, premier models to open platforms they can adapt in-house – and must weigh cost, performance, and trust factors when choosing the right fit.

This week’s developments offer a preview of how AI may evolve in the coming 6–18 months. We can expect a continued pace of rapid improvement: more specialized model tiers for different needs, more multimodal capabilities (like vision and voice) built in by default, and more “autonomous AI” use cases proving themselves in production. The gap between experimental AI and everyday business operations will keep shrinking ([1]), making it imperative for organizations to pilot these tools early and identify high-impact applications.

At the same time, the events of the week underscore the importance of AI governance. With powerful models now capable of unintended behaviors (OpenAI’s Astra being a case in point), and with new regulations like the EU AI Act and California’s SB 942 coming into effect ([2]), companies must build risk management into their AI strategy. That means vetting model providers for compliance features (e.g. transparency and output tracking), setting up strong internal controls (from data governance to human oversight of AI decisions), and ensuring “kill switches” and audit logs are in place for any autonomous AI processes ([3]).

In light of the shifting landscape, C-level leaders should take a fresh look at their AI roadmaps. The dramatic cost reduction in top-tier AI may justify scaling up deployments or pursuing new use cases sooner than anticipated. Conversely, the newfound capabilities of these models – and their integration into tools from Microsoft, Google, and others – means competitors could soon leverage them to accelerate product development, customer service, and more. In short, staying competitive will require rapidly embracing the latest AI tools, but doing so with careful oversight. Leaders would be wise to:

- **Recalculate ROI** for AI projects given that model access is far cheaper and more scalable than even a quarter ago.

- **Run pilot projects** with new models or agent tools on high-impact workflows to assess performance gains and integration challenges early.

- **Double-check compliance and security** by updating AI policies: ensure vendors can meet emerging transparency standards (e.g. content watermarking, audit logs) and implement guardrails to prevent misuse or rogue AI behavior.

Taken together, the capability frontier of AI is advancing at a remarkable rate. Those who can harness the new generation of foundation models stand to gain a competitive edge in productivity and innovation. But realizing that potential, while protecting the business from new risks, will require both urgency and diligence in the months ahead.