Token management for researchers, and why to buy usage before you hire help

What you are paying for

A token is a piece of a word. Models charge for the tokens they read (input) and the tokens they write (output). The context window is everything the model has in front of it at once: your instructions, the files it opened, the conversation so far.

An agent rereads that context on every step. So on a long task the bill is mostly input, and the same instructions and files are paid for again and again unless you do something about it.

Three settings that cut the bill

The waste	The fix	What it saves
Sending the same instructions and reference files on every call	Prompt caching	Anthropic prices cache reads at 0.1 times the base input price
Running a large scoring job in real time	Batch processing	50% off at Anthropic, where most batches finish within an hour. Bedrock prices batch 50% below on-demand
Loading every file “just in case”	Give the agent a manifest and let it open what it needs	Fewer tokens and better answers

The third row matters most, and it is not about money. Anthropic’s documentation says it directly: “As token count grows, accuracy and recall degrade, a phenomenon known as context rot.” A crowded context makes the model worse. Start a fresh session for a new task, keep instructions short, and point the agent at files instead of pasting them.

What a serious result costs

OpenAI reported that the average result in its October 2026 mathematics release used compute equal to roughly three hours of its top-tier thinking. That is the scale at which new results appear: hours of the best model on one problem.

A capped plan stops long before that. The researcher on it sees a tool that summarizes and autocompletes, and concludes that is what the technology is.

Buy usage before you hire help

When a project falls behind, the usual answer is another pair of hands. Compare the two purchases.

	Hiring support	Buying more usage
Time to first result	Weeks to recruit, then training	Minutes
What you must do	Explain the task until someone else understands it	Write the task down precisely once
When the spec is wrong	You find out when the work comes back	You find out in the same hour and fix it
What is left when the project ends	Experience that leaves with the person	Instruction files, checks and pipelines that carry to the next paper
Near a deadline	Fixed hours	As much as you can direct

Two things drive the difference.

The communication gap. Most of the cost of delegating is explaining. You have to specify the task either way. With a person, each round of misunderstanding costs days. With an agent it costs minutes, so you can afford to be wrong three times before lunch.

Skills compound. Every task you run through an agent leaves something reusable: a tested data pull, a checks file, a clearer instruction. Your own skill at directing the tool grows too, and it applies to every later project. Hours bought from someone else do not accumulate that way for you.

This is not an argument against people. Judgment, accountability and the training of the next generation of researchers still need them, and a doctoral student learns the data by handling it. It is an argument about order: give the people you already have the best tools and enough usage first, then see what work is left.

A budget rule

• Put every researcher who writes code on a plan that does not run out during a working day.
• Use caching and batch for anything repeated or large.
• Track cost per finished task, not cost per month. A plan that looks expensive and finishes the analysis is cheaper than one that looks frugal and stalls.