New frontier models break barriers - is your enterprise ready?

While OpenAI is broadening access, Anthropic is taking a more cautious path with its newest breakthrough. Earlier in April, Anthropic unveiled Claude Mythos 5, a frontier model boasting an unprecedented 10 trillion parameters – by far the largest AI model publicly disclosed ([1]). This powerhouse is built for high-stakes, complex tasks; it can perform deep multi-step reasoning with a precision previously reserved for human experts ([2]). In one demonstration, Mythos 5’s cybersecurity skills proved superhuman – the model identified thousands of software vulnerabilities (even uncovering a 27‑year-old security bug in a widely used operating system) that had eluded human engineers ([3]).

Yet, for all its capability, Mythos 5 is not freely available on the market. Citing potential misuse by attackers, Anthropic has restricted access to this model to about 50 organizations under a tightly controlled preview program called Project Glasswing ([4]). These partners – including tech giants and critical infrastructure firms – can use Mythos to proactively probe their systems for weaknesses, but there’s no public API or general release on the horizon ([5]). It’s an extraordinary stance: for the first time, a leading AI company is openly saying it built something “too capable to release” to everyone ([6]). This precedent is prompting the entire industry to rethink how and when to deploy the most powerful AIs.

The strategic implication for enterprises is twofold. First, the cutting edge of AI now comes with serious dual-use concerns – tools like Mythos can defend or disrupt. Business leaders must anticipate that similar capabilities could be used by competitors (or threat actors) to upend markets and security. Indeed, news of Mythos’s abilities triggered real market tremors; cybersecurity stocks fell when details leaked, a sign that investors suddenly view AI’s disruptive power as an immediate reality ([7]). Second, companies not in the inner circle of an AI provider’s “gated” rollout may find themselves at a disadvantage. To avoid strategic blind spots, enterprises should actively participate in industry consortia or cloud partnerships to gain controlled access to frontier AI for positive uses – and simultaneously invest in AI-driven cybersecurity and governance to protect themselves if (and when) equally advanced AI falls into malicious hands.

This week also highlighted that bigger isn’t the only way to better AI – faster and more efficient models are pushing the frontier in performance-per-dollar. Google DeepMind’s latest Gemini 3.1 lineup exemplifies this trend. Its flagship model, Gemini 3.1 Ultra, achieved a record-breaking 94.3% on a rigorous graduate-level reasoning benchmark, outscoring both OpenAI’s GPT-5.4 and Anthropic’s Claude 4.6 ([1]). At the same time, Google introduced a parallel version called Gemini 3.1 Flash‑Lite that sacrifices a small amount of raw capability in exchange for speed – delivering answers about 2.5 times faster and generating 45% more output per second than previous models ([2]). This bifurcation of “max accuracy” versus “low latency” AI systems reflects a growing demand from enterprises: some applications (like research analysis or complex decision support) value the highest reasoning quality, while others (like customer service chats or real-time analytics) prioritize speed and cost-efficiency over absolute perfection ([3]).

Increasingly, AI providers are attacking the cost problem with engineering breakthroughs. Google’s researchers, for example, unveiled a technique called TurboQuant that can compress a model’s memory footprint by a factor of six without hurting performance ([4]). By reducing the massive memory overhead of large language models, methods like this can dramatically cut hardware and cloud costs for running AI – making high-end models more affordable to deploy at scale. And scale is certainly the name of the game: industry estimates suggest the top four tech firms will spend around $600 billion on AI efforts in 2026 alone ([5]). For enterprises, the takeaway is that the frontier is not just getting more powerful – it’s getting cheaper and faster, too. The gap between cutting-edge capability and practical, cost-effective implementation is narrowing. Business leaders should watch for new offerings (like specialized model variants or more efficient chips) that can deliver near-frontier performance at a fraction of the prior cost, potentially enabling broader rollout of AI in their operations.

The final days of April saw what one analyst called an “enterprise tipping point” for AI adoption ([1]). After years of proofs-of-concept, companies are finally moving from experimental AI pilots to production-scale, AI-driven workflows. A key enabler is a wave of new AI agent platforms and integration tools unveiled in the past week. At Google Cloud Next ’26 (April 22), Google launched its Gemini Enterprise Agent Platform, which turns its Vertex AI cloud into a unified stack for building, governing, and optimizing AI-powered processes across the business ([2]). OpenAI, meanwhile, introduced “Workspace Agents” in ChatGPT for its business users ([3]) – allowing organizations to create custom AI agents that can execute multi-step tasks and collaborate across apps on their behalf. Even enterprise software players are jumping in: Snowflake’s new Cortex GPT platform and Infosys’s Topaz AI Fabric (both revealed in April) offer frameworks for orchestrating AI agents within corporate data systems ([4]). These developments collectively aim to eliminate “pilot paralysis” by providing ready-made solutions for deploying AI in complex real-world operations.

Crucially, many of these frontier AI systems are now inherently multimodal – able to understand and generate not just text, but also images, audio, code, and beyond. For instance, Google’s open-source Gemma 4 models can handle language and vision in one package, even running on a smartphone, which opens the door for AI-driven analytics and assistance in fieldwork or on factory floors. Anthropic is similarly integrating its AI into diverse enterprise tools: on April 28, it released “Claude Connectors” to embed its assistant into software like Adobe’s creative suite, Blender 3D modeling, and other design and media applications ([5]). This means AI capabilities can seamlessly plug into professionals’ day-to-day workflows – from generating marketing graphics to reviewing CAD designs – augmenting human creativity and productivity across domains.

As the technology and tooling mature, adoption is poised to explode. Gartner forecasts that by the end of 2026 about 40% of enterprise applications will have some form of built-in AI agent capability, up from less than 5% today ([6]). Similarly, analysts predict roughly 40% of large company roles will soon involve direct collaboration with AI assistants or co-pilots ([7]). The implication is clear: implementing AI-driven automation and decision-support is becoming a competitive standard. Organizations that aggressively embrace these agent and multimodal AI tools – with proper governance – stand to gain a significant efficiency and innovation edge, while those that linger in experimentation risk falling behind as AI becomes core to how business is done.

Recent events underscore a growing strategic split between open and closed AI models. Proprietary model developers are racing ahead with ever-larger systems, but keeping tight control over them – even if that means limited rollout. At the same time, open-source communities and smaller labs are rapidly making AI more accessible. In fact, the very week Anthropic locked down Mythos, China’s Zhipu AI open-sourced GLM‑5.1 (a 744 billion-parameter mixture-of-experts model) that outperformed both OpenAI’s and Anthropic’s previous-generation models on a coding benchmark ([1]). As one industry watcher noted, “the split is no longer about capability. It’s about control” ([2]) – in other words, cutting-edge AI itself is quickly commoditizing, and the real question is who gets access to it.

Google finds itself uniquely playing both sides of this divide. The company released an advanced open‑weight model family (Gemma 4) under a permissive Apache license for anyone to use or modify, even on devices, while simultaneously fielding its top-tier closed model (Gemini 3.1) via its own cloud services ([3]). By supporting open models (which drive community innovation and adoption) and retaining exclusive rights to premium models, Google is hedging its bets and aiming to dominate AI ecosystems regardless of approach. In contrast, Meta – which previously championed open releases – performed an about-face by introducing its first fully proprietary model, Muse Spark, in April ([4]). This “closed” strategy shift comes alongside a pledge to spend up to $135 billion on AI this year ([5]), as Meta seeks to differentiate its products with AI capabilities that rivals can’t easily replicate. These starkly different approaches highlight the strategic choices tech companies face: broad dissemination and ecosystem growth versus tight control and monetization.

Outside the US, open models are emerging as a strategic path to AI self-sufficiency. European startup Mistral has focused on releasing “open-weight” models that are small and efficient enough to run on affordable hardware, letting enterprises fine-tune and deploy advanced AI on their own terms ([6]) ([7]). Likewise, China’s leading labs are prioritizing openness to reduce reliance on foreign tech: the new DeepSeek V4 model launched with public weights and a million-token context window for agent applications ([8]). DeepSeek V4 did not outperform Western rivals on day one, but it sparked a surge in demand for domestic AI chips like Huawei’s Ascend ([9]) – reinforcing China’s push toward “sovereign” AI stacks amid export controls. Even non-traditional players are entering the fray: in a blockbuster move last month, Elon Musk’s SpaceX acquired his AI startup xAI for $250 billion ([10]), creating a $1.25 trillion “galactic” AI conglomerate aimed at integrating frontier models into satellite networks and robotics ([11]). For enterprises, these developments mean more alternatives to Big Tech’s closed ecosystems. Open-source models (and well-funded newcomers) can offer competitive capabilities at lower cost or with fewer restrictions, enabling companies – and even nations – to pursue AI innovations on their own terms.

The breakneck pace of AI advancement shows no signs of slowing. OpenAI’s leadership has hinted that the “surprisingly slow” progress of the past two years is over, and we should expect extremely rapid gains in the medium term ([1]). Indeed, frontier model releases that once came yearly are now arriving in a matter of weeks, each leap bringing new capabilities or efficiencies. For enterprise strategists, this means that agility is paramount – AI roadmaps should be revisited and revised continually as new possibilities emerge.

Over the next 6–18 months, several trends are likely to shape competitive dynamics. First, access to top-tier AI will broaden: with multi-cloud availability, more providers (and integrators like Oracle or Salesforce) will offer advanced models as a service. This will intensify price competition and make AI a ubiquitous utility in enterprise tech stacks – something leaders should leverage when negotiating contracts and designing systems. Second, the integration of multimodal and agentic AI into business processes will mature rapidly. We will see more off-the-shelf AI solutions embedded in everything from customer support platforms to R&D tools, raising the bar for what customers and employees expect in terms of intelligent automation.

Finally, the open vs. closed source dynamic will continue to evolve. We might witness open models achieving parity with today’s best closed models in certain domains – requiring incumbents to innovate faster or risk commoditization of their competitive edge. At the same time, governments and regulators may become more actively involved (especially if another “too-powerful” model like Mythos forces policy intervention). Business leaders should plan for a future where having a robust AI governance framework is as important as having the AI capabilities themselves. In this environment, the winners will be companies that can adapt quickly – adopting new AI innovations, managing AI risks, and striking the right balance between leveraging external AI platforms and cultivating their own AI assets for differentiation. The capability frontier is moving fast, and strategic foresight and flexibility will be the keys to staying ahead.