Every AI subscription bill has a line item nobody explains: the gap between what you actually asked the model to do and what you got charged for. Most of that gap is architecture, not usage — and once you see how context windows and connectors actually work, it's fixable in an afternoon.
The context window is a bill, not a workspace
A large language model doesn't remember your conversation the way a person does. Every time you send a message, the system resends the entire conversation history back through the model — every prior question, every answer, every document you attached — because the model has no persistent memory between requests. That resent history is the context window, and providers charge per token for all of it, not just your new message. A thirty-turn conversation doesn't cost thirty times one message by accident; it costs that much because the system is reprocessing the previous twenty-nine turns every single time.
Pricing has actually gotten more forgiving on raw context size at some providers. Anthropic dropped its long-context surcharge on Claude in March 2026, so a 900,000-token prompt now bills at the same per-token rate as a 9,000-token one. Not every provider has followed that lead: Google's Gemini 3.1 Pro still doubles its input rate once a prompt crosses 200,000 tokens, and several other model families step up their pricing tiers at their own context thresholds. Check the pricing page for the specific model you're using, because "large context window" and "flat-rate large context window" are not the same thing.
Connectors and tools cost money even when you never use them
This is the part almost nobody explains, and it's the biggest hidden cost for anyone running an AI assistant wired into email, calendars, project boards, or a dozen other services. Every connector, plugin, or MCP (Model Context Protocol) tool enabled in a session gets its name, description, and full parameter schema loaded into the model's context on every request — whether that tool gets used that turn or not. Independent measurements put the cost at roughly 550 to 1,400 tokens per tool definition. Connect three moderate-sized integrations — say a code host, a chat tool, and a monitoring service — and around 40 tool definitions can eat 70 percent or more of a 200,000-token context window before a single question gets asked.
One documented case: a developer running ten connected services at roughly fifteen tools each was burning close to 75,000 tokens of pure tool-definition overhead at the start of every conversation, before any real work began. At ten conversations a day, that adds up to well over a thousand dollars a year spent loading capabilities that mostly sat idle.
The fix providers are building toward is on-demand tool loading — the model gets a lightweight index of what's available and pulls the full definition for one tool only right before it's used, instead of loading everything up front. Early implementations report roughly 85 percent reductions in this specific overhead. Until that's the default everywhere, the manual version of the same fix is on you: disconnect or disable integrations you're not using in a given session.
Habits that actually move the number
- Load only what the task needs. Don't run a session with every connector and plugin switched on "just in case." Each one taxes every message, used or not.
- Match the model to the job. Flagship models typically cost five to ten times more per token than a smaller, faster model in the same family. Bulk, templated, or mechanical work — data entry, formatting, first-pass drafts — doesn't need the most expensive model on the menu; save that for judgment calls.
- Practice session hygiene. Start a new conversation for a new task instead of extending one marathon thread. Since every turn resends the full history, a long-running chat is quietly the most expensive way to work.
- Use caching where it's offered. Repeated content — a system prompt, a reference document, a codebase — billed through prompt caching runs roughly 90 percent cheaper on repeat reads than fresh input tokens.
- Batch what isn't urgent. Asynchronous batch processing typically runs at half the price of real-time requests for work that doesn't need an instant answer.
None of this requires understanding how a transformer works. It requires knowing that a connected tool is a recurring line item, that a long conversation is a compounding one, and that the model you pick changes the bill by a multiple, not a percentage. Check what's connected, check what's running, and turn off what isn't earning its keep.
