에이전트 AI가 토큰을 새로운 비즈니스 지표로 만드는 법

Frontier Radar #3: How agentic AI is turning tokens into a business metric Maximilian Schreiner View the LinkedIn Profile of Maximilian Schreiner & Matthias Bastian Jun 8, 2026 Nano Banana Pro prompted by THE DECODER Monthly subscription, open chat, ask a question: that's how generative AI worked until now. Agentic workflows blow up this model. They burn through far more tokens, run autonomously for hours, and make flat rates untenable for providers. At the same time, token prices are splitting along axes of speed, specialization, and economic value. But while costs get more precise, the benefits often stay vague. The result: token usage becomes a stand-in metric for value creation, even though it only measures activity, not outcomes. Six times a year, THE DECODER's editorial team takes an in-depth look at a fundamental AI topic in its "Frontier Radar," as a newsletter and exclusively here on the site for THE DECODER subscribers . Issue #3 covers the emerging token economy of generative AI. Issue #1 looked at the current state of agentic AI. Issue #2 examined the measurable impact of AI on productivity. For a long time, generative AI felt like classic software. Sign up for a monthly plan, open a chat, ask a question, and get an answer. Power users could always see through APIs what individual requests actually cost. That's why many of them went with flat rates, which were much cheaper under heavy use. For most users, though, the costs stayed invisible. Flat rates worked broadly because human usage has natural limits. People type slowly, read answers, take breaks, go to meetings, and clock out. An agent doesn't know those limits. It reads files, calls tools, writes code, checks intermediate results, fixes errors, and tries again. If the user wants, it keeps going until the task is done. There's also the pressure on the provider side: The big AI companies have poured hundreds of billions of dollars into data centers, chips, and model training. Those investments have to pay off, at a scale that flat rates simply can't support. This issue of the Frontier Radar maps out the emerging token economy along these lines. How is billing shifting from subscription to usage? How is the token itself becoming a segmented product? And why is token usage still a poor measure of AI value? Why providers are walking away from flat rates The most visible change is the overhaul of pricing models in response to growing usage. Starting June 1, 2026, GitHub Copilot is gradually moving to a usage-based model with " GitHub AI Credits ." The credits are tied to actual token usage and the API prices of each model. They kick in wherever Copilot does more than just suggest code, mainly in chat, CLI, and agent features. Standard completions stay free of these rules in paid plans. GitHub's reasoning nails the problem: a short chat question used to be treated about the same as an autonomous coding session running for hours. That can't last. Anthropic is also drawing a sharper line between normal use and agentic workflows. Claude Code, Claude Cowork, and Managed Agents turn Claude into a digital worker. Anthropic blamed bottlenecks at Claude Code on peak loads and contexts of up to one million tokens. The older plans fit heavy chat use but not always-on agent workflows. How sharply usage differs between fields shows up in Anthropic's own analysis of its public API : nearly half of all agentic tool calls go to software development, the area that first benefited from agentic models and scaffolding like Claude Code. Customer service, sales, finance, and e-commerce each sit at just a few percent. Simple chat requests still dominate there. That spread will likely widen as agentic workflows mature in office, research, finance, and legal tools. With it, the token bill moves into areas where it isn't yet felt today. Why the token price alone is misleading This development shifts the cost question: As long as AI was mainly used as a chat tool, the price per token could feel like a technical footnote. In agentic workflows, it becomes a business metric. The most obvious mistake in the new token economy is a flat price comparison. GPT-5.5 costs $30 per million output tokens, DeepSeek V4 Pro 87 cents. That says little about actual costs in use. Beyond price per token, what matters is consumption per task. Like with a car, the price of gas alone tells you nothing about what a drive from Berlin to Munich costs. You also have to know the distance and the mileage. A cheap model can get expensive if it needs more tries, fails more often, or requires more cleanup. A pricier model pays off when it gets to the goal with fewer loops and needs less human oversight. Benchmarks and other analyses make this clear. GPT-5.5, for instance, was supposed to offset part of its higher list price with shorter answers. An analysis of real-world usage by OpenRouter still showed cost increases of 49 to 92 percent over its predecessor , depending on input length. Of course, both can rise: the token price and the number of tokens consumed, as with Google's Gemini 3.5 Flash . Here, the token price jumped threefold over the predecessor Gemini 3 Flash. In Artificial Analysis's evaluation, the model also needed more steps in the Intelligence Index run. The result: in that test, it ended up more expensive than Google's current flagship, Gemini 3.1 Pro. Pushing the other way is the price pressure from providers like DeepSeek . Behind the rock-bottom prices is a bet of its own: if you pay only a fraction per token, you can run the same job four or five times and still come out cheaper. As long as the final result holds up, that's attractive. Where it doesn't, rework quickly eats the price advantage. How the token market is splitting by performance class The more the market splits, the less sense it makes to talk about "the" token price. The price per million tokens still matters but only says something within a clear performance class. A fast token in a coding agent, a cheap token in a mass-market app, and a specialized token in security analysis can be billed in similar technical fashion, but they're different economic products. Different model tiers and subscription levels have existed for a while. What's new is that the differentiation now spans more axes: latency, processing mode, context size, agent runtime, specialization, and increasingly the economic value of the output. Providers aren't just selling compute time in token form anymore. They're selling different inference services. The scarcer, faster, or more valuable that service is, the further the price can drift from raw compute costs. Nvidia CEO Jensen Huang spelled this out in two recent interviews. On Dwarkesh Patel's show, he explains why Nvidia recently licensed the inference architecture of startup Groq and folded it into its own CUDA ecosystem. The reason is economic: the value of a token has risen so much that different prices for different token types now make sense. Back in the old days, just a couple of years ago, Tokens were either free or barely expensive. But now you can have different customers, and those customers want different answers. Because the customers make so much money - for example, our software engineers - if I can give them much more responsive Tokens so that they're even more productive than they are today, I would pay for it. Jensen Huang, Nvidia Huang is describing the technical side of this segmentation. Premium inference with lower latency pays off because tokens at the top of the market can command much higher prices. Nvidia talks about expanding the Pareto front: multiple optimal points of price and speed, depending on customer segment. Where the value comes from the possible outcome, there is more segmentation possible. According to The Information, Palo Alto Networks tested Anthropic's security model Mythos to scan its own source code for vulnerabilities. The model reportedly found more than two dozen critical vulnerabilities in a