The Meter Is Running: Using AI in the age of Tokenomics 

| AI

The buffet is closed

For years, AI assistants inside Microsoft’s ecosystem behaved like an all-you-can-eat buffet: pay a flat seat price, use as much as you want. In 2026, Microsoft closed the buffet. Across Copilot, the new Copilot Cowork agent, and the capabilities piloted through the Frontier program, billing now follows the same logic that governs raw API access; you pay for the tokens you consume. This article walks through what changed, contrasts it with the position Anthropic has staked out for Claude, and lays out concrete ways to keep your costs under control. 

What Microsoft actually changed 

Copilot Cowork 

Copilot Cowork, Microsoft’s agentic layer that executes long-running, multi-step work rather than answering a single prompt, reached general availability on July 1, 2026. It bills separately from your Copilot license using a metered unit called a Copilot Credit, priced at $0.01 per credit under pay-as-you-go, where each task’s cost is driven by the models it runs, the context it retrieves, the tools it calls, and the runtime it needs. 

Microsoft groups tasks into rough tiers: light (≈100–300 credits, roughly $1–$3), medium (≈400–700 credits), and heavy (700+ credits, about $7 and up). Cowork still requires a paid Microsoft 365 Copilot seat at $30/user/month on top of the usage-based credits — so it’s a seat-plus-usage model, not a replacement for the seat. Notably, Cowork is disabled by default; no task runs for any user until an admin explicitly enables usage-based billing and sets a spending policy. 

What Anthropic says about charging users 

Anthropic has drawn a deliberate contrast, though it’s worth being precise about what its public position covers. On February 4, 2026, Anthropic published “Claude is a space to think,” a pledge that Claude will remain ad-free. The essay commits to no sponsored links beside conversations, no advertiser influence on responses, and no third-party product placements users didn’t ask for. 

The reasoning is directly relevant to the token-billing debate. Anthropic argues that ad-supported models create an incentive to optimize for engagement — time spent and return visits — which isn’t necessarily aligned with being helpful, since the most useful AI interaction might be a short one that resolves the request without prompting more conversation. In other words, Anthropic’s stated interest is in Claude giving you the efficient answer, not the maximally engaging one. 

This is not the same as “no usage-based pricing.” Anthropic’s API is itself metered per token, and its consumer and enterprise plans are paid subscriptions; the company generates revenue from enterprise contracts and paid subscriptions rather than advertising. The meaningful distinction is about incentive alignment: Microsoft’s move ties cost to token volume, and Anthropic’s pledge is that at least the assistant won’t have a hidden commercial reason to inflate the conversation. For anyone managing a budget, both facts point the same direction. Fewer wasted tokens is the goal, and the design of the tool should not be working against you. 

Using Skills to control token use 

One of the most effective technical levers for controlling token spend is Agent Skills which are the open standard Anthropic released on December 18, 2025. Skills are “organized folders of instructions, scripts, and resources that agents can discover and load dynamically to perform better at specific tasks. Skills extend Claude’s capabilities by packaging your expertise into composable resources for Claude, transforming general-purpose agents into specialized agents that fit your needs.” Anthropic 

The reason Skills matter for cost is a design principle called progressive disclosure. Instead of loading a giant instruction manual into context every session, only each skill’s name and one-line description load at startup costing roughly 30–100 tokens per skill with the full instructions loading only when a task actually triggers the skill, and reference files loading only during execution. Anthropic 

The token savings are concrete. In one measured example, the startup descriptions for ten installed skills required only 773 tokens, whereas eagerly loading the full instruction bodies would have required about 13,900 tokens, and loading bodies plus scripts about 65,900 tokens. That is close to a 20x reduction in baseline context cost for the same capability set. Because tokens are the billing unit, that reduction flows straight to your invoice. 

Practical ways to use Skills to control spend: 

  • Push rarely needed detail into reference files. Anthropic’s own guidance is to use progressive disclosure aggressively: if information is needed only ~20% of the time, it belongs in a reference file, not the main body, so you don’t pay for it on every run. 
  • Write tight trigger descriptions. A precise name/description means the skill fires only when genuinely relevant, avoiding wasted loads. 
  • Encode repeated workflows once. If you retype the same report format, review checklist, or analysis pipeline across conversations, a skill turns that into a reusable, deterministic asset instead of re-prompting (and re-paying) every time. 

Understand your workflow: put AI where it adds value 

The single biggest cost lever isn’t a setting but it’s knowing where in your workflow AI actually earns its keep. A useful exercise is to map your workflow into three buckets: 

  1. High-value, AI-appropriate steps. Genuinely hard reasoning, synthesis across many documents, first drafts of complex artifacts, multi-file refactors. Here the token cost buys real leverage so you can spend confidently. 
  1. Low-value or deterministic steps. Formatting, boilerplate, repetitive transformations, simple lookups. These are cheap or free with a script, a template, or a lightweight model. Sending them to a frontier agent is pure waste. 
  1. Steps where AI adds risk, not value. Anything where you’d have to carefully verify the output anyway, or where a wrong answer is expensive. The token cost plus your review time may exceed just doing it directly. 

Once you can see the map, the cost strategy writes itself: reserve the expensive models for bucket one, automate bucket two deterministically, and keep AI out of bucket three. Microsoft effectively built this thinking into Cowork via its per-task model picker — the recommendation to pick a lighter model where it fits, keep tasks tightly scoped, and watch for lower-cost models like the forthcoming Cowork 1 is a direct application of workflow-aware spending. 

The point isn’t to minimize AI use. It’s to spend tokens where they convert into value and stop spending them where they don’t. 

Other ways to control costs 

Beyond Skills and workflow mapping, a handful of governance and habit changes make a large difference: 

  • Set hard spending limits before rollout, not after. Microsoft Cowork’s Cost Management dashboard lets admins set spending limits and alerts per tenant, group, and user. Configure caps and alerts before you enable usage, so consumption can’t surprise you. 
  • Choose the right model for each task. Model choice is one of the biggest cost levers in every one of these products. Route quick questions to lightweight models and reserve frontier models for work that needs them. 
  • Match your pricing plan to your usage pattern. Pay-as-you-go suits spiky, unpredictable demand with no commitment; prepaid volume plans (Microsoft’s P3 pre-purchase, for example) suit steady usage and buy budget certainty at a better effective rate. Estimate before you commit with Microsoft’s Copilot Credit Estimator. 
  • Scope tasks tightly. Vague, open-ended prompts invite bloated, multi-iteration agent runs. A specific request with clear boundaries consumes far fewer tokens than “figure this out and keep going.” 
  • Avoid redundant iteration. Much of the runaway spend developers reported came from repeated, bloated agent loops. Review and correct direction early rather than letting an agent churn. 
  • Use included allotments first. Each plan bundles a monthly credit allowance. Track how much of its real work consumes before buying overages, and use the preview/projected-cost tooling Microsoft rolled out ahead of the transitions. 
  • Monitor continuously. Analysts advise closely watching cost impact as pricing shifts to PAYG to decide whether deployment or governance changes are needed to stay within budget. Treat token spend as an operational metric, not an afterthought. 

The bottom line 

Microsoft’s shift to token-based billing for Copilot Cowork capabilities isn’t an aberration; it’s the industry acknowledging that agentic AI is too expensive to give away flat-rate. That makes efficiency your responsibility. The good news is that the same practices that cut cost such as leaner context via Skills, workflow-aware model selection, tight scoping, and hard budget caps also tend to produce sharper, faster results. Whether your assistant’s incentives are aligned with brevity or not, the winning move is the same: spend tokens where they create value, and nowhere else. 

References