Bigger, Smarter, Cheaper: AI’s New Frontiers Rewrite the Enterprise Playbook

In the past week, AI leaders OpenAI and Anthropic rolled out significant upgrades to their foundation models, underscoring the blistering pace of progress at the capability frontier. OpenAI released GPT-5.6 in preview, including specialized variants (Sol, Terra, and Luna) aimed at complex, multi-step tasks ([1]). Anthropic debuted Claude Sonnet 5 – an *agentic* upgrade to its Claude model that arrived just a few weeks after its previous version – and touted it as the lab’s most autonomous AI yet ([2]). Industry trackers note that major AI models are now launching roughly every two days in 2026 ([3])，highlighting how rapidly state of the art is being redefined.

Notably, *agentic* capabilities – the ability for AIs to plan, use tools, and perform tasks autonomously – emerged as a unifying theme in these releases. OpenAI’s GPT-5.6 Sol enables work to be split among multiple cooperating sub-agents for extended tasks ([4]). Google’s Gemini 3.5 Flash, which launched in May, was similarly described as shifting from a chatbot to an AI that can "plan, build, and iterate on real work" with minimal human input ([5]). With Anthropic’s Claude Sonnet 5 also emphasizing tool use and autonomous workflows ([6]), it’s clear that strategic focus has shifted to making top models more *agentic*.

Another breakthrough this week came not from model architecture, but from silicon. OpenAI unveiled Jalapeño, its first custom-built AI inference chip developed with Broadcom ([1]). Early tests show this specialized processor delivers "significantly better performance-per-watt" than even the best general-purpose AI hardware ([2]). OpenAI credits its full-stack approach – from model design to chip architecture – with making its AI faster and more cost-efficient ([3]).

Just days later, rival Anthropic revealed it’s exploring a similar path. According to media reports, Anthropic is in early talks with Samsung to develop its own AI chips ([4]). The goal: create custom silicon that lowers the cost of running large-scale AI models and reduces reliance on Nvidia’s dominant GPUs ([5]). (Both Amazon and Google have likewise built specialized AI processors – AWS’s Inferentia/Trainium and Google’s TPUs – to achieve better price-performance in the cloud ([6]).)

For enterprises, the takeaway is that the cost and speed of AI are poised to improve significantly. Training cutting-edge models has historically required tens of thousands of high-end GPUs and enormous cloud resources – a budget only tech giants could afford – but these new chip initiatives could bend the AI cost curve downward. OpenAI notes that even small reductions in inference cost can greatly improve the bottom line ([7]), which means advanced AI services will become more affordable and accessible. However, if the most powerful models become tied to proprietary hardware, businesses must plan their partnerships carefully to avoid being locked into a single ecosystem as the AI stack becomes more vertically integrated.

Once, only companies like OpenAI, Google, or Anthropic could offer the most advanced AI models – but that reality is changing. Over the last 18 months, the performance gap between proprietary (closed) models and open-source alternatives has rapidly compressed. The Stanford AI Index 2025 report found that by early 2026, the difference in accuracy on key knowledge benchmarks had shrunk from a 17.5-point lead for closed models in late 2023 to effectively zero ([1]).

Multiple independent efforts have driven this convergence. Open-model pioneers in both academia and industry – from Chinese tech labs to startups – have matched or surpassed traditional leaders on various tasks. For example, recent open-model achievements include:

- **Alibaba’s Qwen 3.5** – scored 88.4% on a challenging graduate-level Q&A benchmark, outperforming every closed model except the largest frontier systems ([2]).

- **Moonshot AI’s Kimi K2.5** – a 1-trillion-parameter open model – hit 99.0% on the HumanEval coding test ([3]), surpassing even the best proprietary code models in accuracy.

These milestones demonstrate that open-source AI can now deliver top-tier results in areas from general knowledge to software development. For enterprise strategy, this rise of open foundation models creates both opportunity and complexity. On one hand, companies have access to cutting-edge AI capabilities without depending solely on a few vendors. Open models (often released under permissive licenses like Apache 2.0) can be self-hosted and fine-tuned on proprietary data, appealing to organizations that value control, customization, and cost savings. On the other hand, closed-source providers still offer advantages in certain domains – for example, the very latest reasoning or *agent* features, or dedicated support and compliance certifications – and they are quickly lowering prices and improving usability in response to open-source competition. Leadership will need to continually weigh the benefits of open models versus proprietary services, and may increasingly adopt hybrid strategies that combine the best of both.

Today’s frontier models are not only more powerful – they’re more versatile. The latest generation of foundation AIs can handle multiple data modalities in one system, ingesting and analyzing text, images, audio, and even video concurrently ([1]). This multimodal ability opens up a host of enterprise applications – from generating rich marketing content (combining text and graphics) to analyzing visual documents and video archives for insights – all with a single AI assistant instead of separate specialized tools.

Equally important, these models have dramatically expanded “memory.” Where early GPT-3 managed about 4,000 tokens (a few pages of text) at a time, new models boast context windows of hundreds of thousands to over a million tokens ([2]). In practical terms, an AI can now absorb and reason over entire corporate reports, large datasets, or lengthy legal contracts in one go. Long-context AIs can identify patterns and draw conclusions from a company’s troves of information without needing to split or summarize the input, enabling more comprehensive analysis and holistic decision support.

Beyond understanding content, next-gen AI can take action. Advanced *agent* capabilities allow models to execute code, call external APIs, search databases, or control software tools as part of their responses ([3]). Instead of just producing an answer, an AI might autonomously perform tasks – for example, retrieving real-time data, running an internal simulation, or composing and sending emails. This turns AI from a passive information provider into an active digital assistant or co-worker capable of carrying out multi-step projects.

However, these new abilities also come with challenges. Fully autonomous AI agents are still in their infancy, and even tech optimists acknowledge that progress has not been as quick as initially hoped ([4]). Early adopters report that while AI agents can initiate and execute tasks, human oversight and clear guardrails remain essential to ensure quality, security, and compliance. As *agentic* AI becomes more common, enterprises will need robust governance and risk management—but those that get it right stand to unlock unprecedented efficiency and innovation.

For business leaders, the furious rate of AI advancement means the competitive landscape can shift in a matter of months. In the coming 6–18 months, further leaps in capability are likely as companies pursue both sheer scale and new techniques. We may even see early versions of truly continuous learning (AI systems that learn on the fly from new data) and greatly improved long-term memory beyond today’s huge context windows – breakthroughs that DeepMind’s CEO Demis Hassabis calls critical for more adaptive, personalized AI ([1]). At the same time, expect ever-larger and more sophisticated models from the major providers, as well as new open-source entrants pushing the envelope in specialized domains.

In this environment, strategic agility is key. Organizations at the forefront are already employing a portfolio of models rather than relying on a single AI for every task. They route each request to the AI that best meets its requirements – for example, using a top-tier model for complex analysis but a smaller open-source model for routine summaries – all coordinated by intelligent routing systems ([2]). This multi-model strategy can dramatically lower costs (some reports suggest savings of up to 80%) without sacrificing performance, and we expect more tools to emerge that support such smart orchestration of AI services.

Finally, executives should continually revisit their “build vs. buy” decisions as AI capabilities and economics evolve. With open models reaching parity in many areas, it’s becoming feasible to fine-tune or customize them in-house for specific needs – potentially saving on vendor fees and keeping sensitive data in-house. On the other hand, proprietary AI services still offer advantages in ease of deployment, security assurances, and earliest access to cutting-edge features. The likely best approach for many enterprises will be a hybrid one: combine proprietary and open-source AI solutions, and remain ready to switch components as capabilities and costs shift. Above all, staying at the capability frontier will require proactive experimentation, workforce upskilling in AI, and close monitoring of the AI landscape. The winners will be those who act now to integrate these rapidly evolving AI capabilities into their strategic roadmap, ensuring they are not just reacting to change but leading it.