AI progress isn’t just about making smarter models anymore. The real measure of advancement is how much compute we can meaningfully consume at one time, coupled with the dramatic improvements in intelligence per unit compute. This isn’t a subtle shift; it’s a fundamental reorientation. What was theoretically possible a year ago—like long-running AI agents—was prohibitively expensive and slow. Now, with much more efficient models that can reason better, these agents can consume more compute at once because they remain coherent longer. This sustained coherence is now a primary driver of progress.
Consider ChatGPT Agent. The models were smart enough to run something like that maybe a year ago, but it would have been extraordinarily expensive and just not viable. It would have taken far too long. Now, we have models that can maintain coherence over extended periods, enabling them to consume more compute without falling off. This is a far more significant advancement than raw intelligence gains alone, though those have been massive too.
The Coherence Breakthrough: Sustained Progress Over Time
The ability for AI models to maintain coherence and productivity over extended periods is now the primary marker of practical progress. Early agents like Deep Research could sustain tasks for tens of minutes. Today’s frontier models, however, can autonomously operate for hours, executing complex workflows and research tasks that were previously out of reach for AI.
Claude 3.5 Sonnet, for example, has impressive raw coding capability. Its progress in that area is substantial. But the more profound advancement is its ability to remain coherent over longer time periods. When older models, like Claude 3.5 Sonnet, were playing Pokémon, they often stalled after a certain amount of progress because they couldn’t operate in a long-running context effectively. Current models have largely overcome this, demonstrating a significant leap in agentic progress.
The task length frontier has doubled every seven months, with the latest models handling tasks approaching an hour in human-equivalent time.
These great applications—Manus, ChatGPT Agent, and powerful coding agents like Claude Code—can keep running and making progress toward a task for hours. This represents a primary vector of progress. However, it also keeps costs high because at any given time, 90% of demand is on the most powerful frontier model available.
The Cost Paradox: Affordable Intelligence, Expensive Frontier
Here’s where the economics of AI get interesting. For a particular level of intelligence, the cost continues to drop dramatically. If you want a model that is about as smart as GPT-4, you can get that for dirt cheap now. GPT-5 Nano, for instance, costs $0.05 input and $0.40 output per million tokens. It’s significantly smarter than GPT-4 in almost every regard and way more useful in general.
The cost per unit intelligence has plummeted. But what people actually use—the most frontier models—consumes tons of compute, and that cost is only increasing. This creates a two-tier system where basic intelligence is becoming commoditized, while cutting-edge capabilities command premium prices.
| Model | Intelligence Index | Run Cost | Specialization |
|---|---|---|---|
| Grok-4 | High | $1658 | Long agentic tasks |
| GPT-5 (Thinking) | High | $823 | World knowledge, reasoning |
| GPT-5 Nano | 56 | $41 | General tasks, affordable |
Current frontier models showcase the high-cost, high-capability paradigm while smaller models offer GPT-4+ intelligence at dramatically lower costs.
This pricing structure reflects a clear market reality: the most powerful models are always relatively stable or increasing in cost because they are the only ones capable of sustained, agentic operation on complex, hours-long tasks. The infrastructure and compute requirements for these frontier models remain substantial, justifying their premium.
Frontier Model Landscape: The Current Champions and Their Strengths
The current frontier is dominated by a handful of systems that exemplify this high-cost, high-capability paradigm. OpenAI’s GPT-5 series, especially the ‘Thinking’ variant, stands out as the ‘King’ for its world knowledge and raw intelligence. Its primary competitor for the top spot is xAI’s Grok-4, the ‘Grandmaster’ model, which excels at the very long-running agentic tasks that define the latest wave of progress.
Other major players in this top tier include Google’s Gemini 2.5 Pro, a ‘Sorcerer’ with a massive context window and multimodal strengths, and Anthropic’s Claude Opus 4.1, a ‘Goldsmith’ specialized for high-end coding and creative work. These are the models that command premium prices because they possess the advanced coherence and reasoning necessary to meaningfully consume vast amounts of compute for complex, hours-long tasks.
Claude 3.5 Sonnet’s Coding Prowess
Claude 3.5 Sonnet deserves special mention for its industry-leading coding capabilities. In internal evaluations, it solved 64% of agentic coding problems, a significant jump from 38% for its predecessor. On SWE-bench Verified, it improved from 33.4% to 49.0%, achieving the highest score among all publicly available models. This isn’t just about raw coding skill; it’s about the model’s ability to independently write, edit, and execute code, troubleshoot, and plan multi-step tasks without human intervention.
Real-world users from companies like GitLab, Cognition, and The Browser Company report substantial improvements in coding, planning, and automation, with no added latency. The ability to autonomously fix bugs, write tests, and update codebases with minimal human input is now a reality. This level of autonomous coding is a direct result of improved agentic coherence and the ability to process more compute effectively over time.
The Compute Scaling Reality and Its Implications
The amount of compute used for training and running frontier models continues to grow exponentially, with training compute expanding at a rate of about 4x per year. This scaling has outpaced other technological directions and is a major driver of the rapid improvement in AI agent capabilities. This isn’t just about bigger models; it’s about enabling them to do more useful work for longer periods.
This exponential growth in compute is projected to continue through 2030. However, this trajectory depends on infrastructure investments keeping pace with critical challenges in power, chip manufacturing, data, and latency. The ‘task length frontier’—how long a model can reliably work on a task—has doubled every seven months, with the latest models able to handle tasks approaching an hour in human-equivalent time. This is a crucial metric for practical AI applications.
Real-World Performance Benchmarks and User Feedback
The benchmarks tell a clear story about this agentic progress. Claude 3.5 Sonnet’s performance improvements aren’t theoretical; they translate to real productivity gains. Its ability to remain coherent and productive over extended coding sessions has made it a favorite among developers for complex, multi-step programming tasks.
Companies are seeing tangible benefits. The ability to hand off a complex task and have the AI work on it for hours without intervention is becoming a reality rather than a mere promise. This is a direct outcome of the focus on agentic coherence over raw intelligence alone.
The Infrastructure Challenge and Market Dynamics
This progress comes with infrastructure challenges. The 90% demand concentration on frontier models creates supply constraints and keeps prices high for cutting-edge applications. Meanwhile, the dramatic cost reduction for achieving specific intelligence levels democratizes access to AI capabilities that were recently considered advanced.
The dichotomy creates interesting market dynamics. Startups and individual developers can access powerful AI capabilities at low cost, while enterprises requiring the absolute cutting edge pay premium prices. This two-tier system is likely to persist as the gap between frontier capabilities and commodity intelligence continues to widen. It’s a clear signal that the market values sustained, coherent work over raw, fleeting intelligence.
Looking Forward: Compute as the Limiting Factor, Not Just Intelligence
The trend suggests that future progress will continue to be gated by our ability to meaningfully consume compute rather than by algorithmic breakthroughs alone. Models are becoming more efficient at converting compute into useful work, which enables them to tackle longer, more complex tasks without losing coherence. This has profound implications for how we prioritize AI development.
Raw intelligence improvements still matter, but the ability to maintain coherence and make progress over extended periods might be more practically valuable for many applications. The models that can work autonomously for hours are the ones commanding premium prices and driving real productivity gains. They are literally doing more work.
The infrastructure requirements for this continued progress are substantial. Power, chip manufacturing, data availability, and latency all become critical constraints as compute requirements continue their exponential growth. But if these challenges can be met, the trajectory suggests we will see continued improvements in agentic capabilities that make AI assistants increasingly practical for complex, real-world tasks.
The current moment represents an inflection point where AI agents are transitioning from promising demos to practical tools that can genuinely augment human productivity. The combination of improved efficiency and sustained coherence has made this possible, even as it creates new challenges around cost and infrastructure scaling. The future of practical AI is less about pure brainpower and more about endurance and efficient execution.
