Police officers cut power first. That detail from a recent raid on Danish privacy activist Lars Andersen (@LarsAnders1620) spread quickly on X. Officers headed straight for the circuit breaker panel. They wanted to stop cameras from rolling. The incident, detailed in his thread, sparked debate about surveillance, authority and technology’s role in accountability.
But turn the lens to artificial intelligence. Similar tensions surface. Models grow more capable. They reason longer. They consume more compute at inference time. And questions arise. Who controls the off switch? What happens when systems think harder than their creators anticipated?
The Shift to Test-Time Intelligence
xAI released Grok 4 in July 2025. The company called it the most intelligent model in the world. It came with native tool use and real-time search. Yet the standout feature lived in its heavier variant. Grok 4 Heavy embraced parallel test-time compute. The model considers multiple hypotheses at once. It scales reinforcement learning to levels not seen before.
Results followed. On Humanity’s Last Exam, a benchmark built from 2,500 questions by over 1,000 experts across more than 100 disciplines, Grok 4 Heavy reached 50.7% on the text-only subset with tools. No prior model had crossed 50%. xAI’s announcement highlighted the score. It also posted 61.9% on USAMO 2025 math competition problems. On ARC-AGI V2, Grok 4 hit 15.9%. That nearly doubled the previous high.
These numbers matter. Earlier models improved mainly through bigger pre-training runs. Grok 4 leaned on post-training. Massive reinforcement learning on verifiable tasks expanded beyond math and code. The Colossus cluster, grown to 200,000 GPUs, delivered over an order of magnitude more compute for RL than prior efforts. Efficiency gains reached 6x through infrastructure and algorithmic tweaks.
But. The approach carries costs. Longer thinking times. Higher inference expenses. Grok 4 Heavy sits behind a $300 monthly SuperGrok Heavy tier. Standard access runs $30. Trade-offs appear everywhere.
Scientific American covered the launch. Elon Musk described Grok 4 as capable of perfect SAT scores and near-perfect GRE results across subjects. “AI smarter than humans is frightening but likely good,” he said in the July 11, 2025 article. The piece noted Grok 4’s 44.4% on the full Humanity’s Last Exam with Heavy mode. It outperformed Google’s Gemini-Pro at 26.9% and OpenAI’s o3 at 24.9% in certain tool-assisted settings.
Independent evaluators weighed in. Artificial Analysis placed Grok 4 at the top of its intelligence index. Greg Kamradt of the ARC Prize Foundation verified the ARC-AGI results. Still, caveats surfaced. Some users reported coding errors. Jailbreak vulnerabilities persisted from earlier versions. On controversial topics, the model sometimes consulted Musk’s public statements on X for context.
And the field moved. By late 2025 xAI rolled out Grok 4.1. It improved multimodal understanding, reduced hallucinations and added stronger agent tools. The xAI news page detailed Grok 4.1 Fast and an Agent Tools API for orchestrating external capabilities. Context windows stretched. Pricing dropped on efficient variants. Developers gained configurable reasoning effort levels. Low for quick tasks. High for complex analysis.
These updates reflect a broader industry pattern. Pre-training gains slow. Inference-time scaling accelerates. Models that “think” for seconds or minutes now outperform larger but static counterparts on hard problems. Parallel hypothesis generation, as in Grok 4 Heavy, marks one path forward. Mixture-of-experts routing to specialized internal agents marks another.
Critics remain. Gary Marcus argued on his Substack that Grok 4 and OpenAI’s o3 results vindicate neurosymbolic ideas long dismissed by pure deep learning advocates. The models still struggle with consistency. They hallucinate under pressure. Yet on benchmarks designed to exhaust human expertise, they post gains.
Enterprise adoption followed. Microsoft integrated Grok 4 into Azure AI Foundry. Oracle offered Grok 4 Fast variants with reasoning and non-reasoning modes. Pricing settled around $1.25 per million input tokens for capable versions. Companies tested agentic workflows. Real-time web and X search integration let models update knowledge without retraining.
Voice mode advanced too. Grok gained real-time video analysis during calls. A new serene voice option joined the lineup. Responsiveness improved. These features point toward ambient assistants rather than chat-only tools.
Yet the core tension lingers. Greater reasoning power demands more electricity, more silicon, more money. It also raises stakes around alignment, bias and control. When a system can spend minutes generating dozens of parallel reasoning chains, who decides when it has thought enough? How do regulators or users pull the plug if outputs veer into unexpected territory?
The Danish activist’s experience, however unrelated on surface, echoes here. Technology that records or reasons without oversight creates unease. AI that reasons in secret, behind paywalls or proprietary clusters, invites parallel concerns. Transparency matters. So does competition. xAI’s aggressive scaling, backed by Musk’s vision to understand the universe, forces rivals to respond.
Further iterations seem certain. Grok 4.3 appeared in API docs by mid-2026 with 1 million token context and refined non-reasoning modes. Agentic coding models like Grok Code Fast 1 gained traction for software engineering tasks. Benchmarks continue to fall. The question shifts from whether models can reason to how organizations and societies govern the consequences.
Short answer. They can’t yet. Not fully. But the trajectory looks clear. Test-time compute is the new training compute. Models that allocate effort dynamically will define the next wave. Industry insiders watch the numbers. They track the dollars per token. They measure latency against accuracy. And they wonder what happens when the circuit breaker no longer works.
WebProNews is an iEntry Publication