The era of tokenmaxxing — burning tokens to create the appearance of productivity — is receding as providers and users confront real costs. A clear turning point came when GitHub Copilot changed its pricing: instead of a monthly fee with unlimited use, Copilot moved to a monthly plan that provides a limited number of credits. One credit equals US$0.01 and is used to pay the AI provider of the user’s choice; when credits run out, users must upgrade or purchase additional credits.
Why now? Two converging large trends
The shift reflects two major industry trends. First is blitzscaling: starting with OpenAI, many frontier providers prioritized rapid user growth over immediate profitability. Giving services away or underpricing them attracted users in the short term, with the expectation that revenue streams (end-user fees, data sales, advertising) would be found later. That approach tends toward eventual revaluation and stricter pricing.
Second, token consumption surged in late 2025. The arrival of reasoning models increased token counts because these models maintain internal dialogues — reasoning tokens — while solving problems. Reasoning tokens are often more numerous than the input prompt and the final response; they are frequently billed as output tokens and cost more than input tokens.
Agents further multiplied token consumption. As Simon Willison quoted Anthropic’s Hannah Moran in May 2025: “Agents are models using tools in a loop.” An agent loop reads current data, thinks about its meaning, chooses an action, executes it, checks the result, and repeats. Watching systems like Claude Code or OpenClaw shows how a single request can spawn multiple model calls, each using hundreds or thousands of tokens. Agent invocations can also include the task’s accumulated context and relevant documents, so reasoning tokens plus agent activity can increase token use by orders of magnitude.
Price and capacity: the limits of loss-leader growth
Higher token usage might be justifiable if it led to proportionally better outcomes, but blitzscalers’ willingness to operate at a loss is not unlimited. Even if user numbers grow, computation — and thus cost — per user rises as agents are used more widely. That combination has driven pricing changes:
- Microsoft/GitHub moved Copilot to token credits so Copilot customers no longer rely on the provider to absorb costs indefinitely.
- Anthropic and OpenAI have priced more capable models significantly higher: Claude Fable is about twice the price of Opus 4.8, and GPT 5.5 costs roughly twice as much per million tokens as GPT 5.4.
Anthropic’s planned move to token-based billing for its agent SDK (currently on hold) is another sign that cheap AI days are ending. Yet higher-capability models can also increase token use, so upgrading models does not necessarily reduce total cost.
Capacity constraints are another major factor. Big data centers have been discussed in the media, but the centers themselves — and the electrical infrastructure (transmission lines, generators) needed to power them — are not trivial to build and require investments that AI companies do not typically manage. Anthropic has blamed outages on capacity and responded by leasing unused data-center capacity from SpaceX. When demand outstrips available capacity, providers can respond by raising prices and focusing on customers who can afford higher fees. Managers, accountants, and independent developers are noticing these shifts.
Accountability and token optimization
Pressure on token prices naturally leads to accountability and optimization. As Bennie Haelen argues in “The Subsidy Ended: What Tool-Using Agents Actually Cost,” better governance comes from building an observability layer that reveals exactly what agents and models do. A well-designed observability layer shows whether data sent to models grows with each invocation, whether the right tools are used, whether tools are called repeatedly, and other metrics that indicate agent efficiency.
Understanding which models handle requests is another part of accountability. Reasoning models range from expensive, high-performance models like Claude Fable or Opus 4.8 to models such as Gemma 4 26B that can run on a well-equipped laptop, and even smaller local models. Most requests do not need the highest-performing, most expensive model, so agents can decide which model is appropriate for each request. Fable can delegate to less costly Anthropic models, and other frontier providers are expected to follow as models gain agent capabilities. There is also a vibrant ecosystem of open models that run locally; Vicki Boykis notes that local models now often perform nearly as well as frontier models. Tools like OpenRouter provide model-independent routing to different models — including local open models — and can be integrated with agents such as OpenClaw, Claude Code, Cursor, or Codex for intelligent routing.
Conclusions
Tokenmaxxing is on the decline. Vestiges will remain for a while — some developers and managers will still push high-consumption approaches — but responsible token spending is becoming the norm, whether costs are borne personally or by a company. As per-token charges rise, token optimization and observability will grow in importance, and such charge increases seem likely to continue.
