Token Efficiency: Spending an AI Context Budget Wisely

What Does It Mean to Treat Context as a Budget?

Every token in a model's context window gets processed and billed on every single call in that session. Treating context as a budget means deciding on purpose what earns a seat in that window, instead of loading everything you have.

A context window is not free memory. It is paid input, processed fresh on every request, and it does not get cheaper just because the window got bigger. Production context windows now run close to a million tokens on both major platforms, and the per-token price stays the same whether you use 9,000 tokens or 900,000 of it. A bigger window removes a ceiling. It does not remove the bill.

It also does not remove the reliability cost. Our report on context engineering covers the research showing models attend worse to the middle of a long context, so a bloated window is not just expensive, it is often worse at the job. Budgeting context is therefore two problems solved by the same discipline: pay for less, and get a more reliable answer from what you did pay for.

How Does Progressive Disclosure Cut Wasted Tokens?

Progressive disclosure keeps a short router file, roughly 50 to 200 lines, in every context, and moves the detail into topic files a model opens only when the task calls for it. A giant instructions file taxes every call whether the rule applies or not.

The natural instinct after every agent mistake is to add another rule to the main instructions file. Do that enough times and the file hits four or five hundred lines, and now every single task, no matter how small, pays to load rules that do not apply to it. That is a real cost twice over: the tokens themselves, and the attention the model has left over for the actual work.

The fix is a router, not an encyclopedia. A short entry file carries the project overview, the hard constraints, and links to topic files with a one-line description of when each one applies. A task that touches the database loads the database rules. A task that does not, doesn't. The router stays small and cheap on every call; the detail only gets paid for when it is relevant. It is the same idea our harness engineering report covers as the state layer of a reliable agent system, applied specifically to the token bill.

Why Does Delegating Research to a Subagent Save the Main Thread's Context?

A subagent opens its own disposable context window, does the messy exploration, and hands back a short summary. The main thread never pays to hold the search that produced the answer, only the finished answer itself.

Fanning work out to subagents is not automatically cheaper. Anthropic's own engineering team, writing up the multi-agent research system behind Claude's Research feature, reported that agents typically use about 4 times more tokens than a plain chat, and multi-agent systems use about 15 times more, across the whole system. That is the honest number, and it matters: multi-agent fan-out spends more tokens in total, not fewer.

What it buys is a different shape of spending. The main thread's context stays small and focused because the exploration, the dead ends, and the tool output that did not pan out all happened in a subagent's window that gets thrown away when the subtask ends. The parent agent's context grows by one clean result, not by everything it took to find that result. Anthropic's guidance is specific about when this trade is worth it: tasks with heavy parallelization, information that exceeds a single context window, and enough value on the line to justify the extra spend. Our own field note on multi-agent orchestration covers the fan-out pattern in full.

What Do Anthropic and OpenAI Actually Charge Per Token?

Published mid-2026 rates run from about $1 to $50 per million tokens depending on tier, with output priced several times higher than input on every platform. Which model you route a task to is still the single biggest lever on a token budget.

The model tier you pick still swings your bill more than any caching or delegation trick. Here is what the two major platforms publish for standard API pricing as of July 2026:

Two things worth naming from these tables. First, output tokens cost 5 to 6 times more than input tokens on both platforms, so a verbose answer is a bigger lever than a verbose prompt. Second, Anthropic's newest tokenizer, used by Fable 5, Mythos 5, and current Opus and Sonnet models, produces roughly 30 percent more tokens for the same text than the older tokenizer. A per-token price cut can be partly offset by a token count that quietly grew.

How Much Does Prompt Caching Actually Save?

A cache hit costs about 10 percent of the standard input rate on both platforms, so a stable system prompt, tool schema, or reference document you reuse across calls gets nearly free after the first request pays to store it.

Caching targets a specific shape of prompt: a fixed prefix that repeats across calls, followed by a variable tail that changes every time. A system prompt, a set of tool definitions, and a reference document are classic fixed prefixes. The user's latest message is the variable tail. Anthropic prices a cache write at 1.25 times the base input rate for a 5-minute cache or 2 times for a 1-hour cache, and a cache read at 0.1 times, a 90 percent discount on the repeated tokens. OpenAI's caching works the same way: cache writes cost roughly 1.25 times the uncached rate, and cache reads get a 90 percent discount, with a 30-minute minimum cache life.

That math compounds. Combine prompt caching with batch processing, which both platforms discount by roughly 50 percent for non-time-sensitive work, and the effective rate on repeated, batchable tokens lands around 5 percent of the sticker price. The catch is real: caching only pays off if the prefix genuinely repeats. A prompt that changes shape on every call gets none of the discount and eats the write premium for nothing.

Why Does Cross-Session State Also Save Tokens?

A session that reconstructs project context by re-reading a codebase from scratch burns thousands of tokens rediscovering facts a prior session already knew. A short progress file a new session reads first turns that rediscovery into a three-minute read.

Context windows end. Sessions crash, hit their limit, or simply close. If nothing was written down, the next session pays the full re-exploration cost: reading files, re-deriving decisions, sometimes re-litigating a choice that was already settled. The fix is the same one we install on every multi-session build in our harness engineering practice: a progress file that carries current state, and a decisions file that carries the reasoning nobody wants to redo.

There is a related cost that shows up near the end of a long session, not just at the start of a new one. Anthropic's own engineering team has documented that models can rush and skip verification steps as a context window fills up, a pattern they call context anxiety. After moving to a model that stopped showing that behavior, they dropped scripted context resets from their harness entirely and let the agent run as one continuous session. Fewer resets means fewer cold re-derivations, which means fewer tokens spent rebuilding a state that was never lost in the first place.

When Does a Bigger Model Actually Beat Raw Efficiency?

When a task genuinely needs deep, long-horizon reasoning: hard debugging, architecture review, ambiguous multi-step planning. Efficiency wins almost everywhere else, because most failures trace back to a missing router file, not a missing IQ point.

None of this argues against ever paying for a bigger model. It argues against paying for one by default. Our context engineering report covers a pattern we see constantly in client work: the same model fails a task, the context gets fixed, and the same model completes it cleanly with no upgrade at all. Prove the context was right before you pay more for capability.

In practice we route by task shape. Mechanical, well-specified work goes to the cheapest tier that can do it. The bulk of production work, the actual workhorse volume, goes to a mid tier. The top tier gets reserved for genuinely hard, long-horizon reasoning: the kind of review and judgment work we cover in our field comparison of Claude Fable 5 and GPT-5.6 Sol. That routing decision, made deliberately per task instead of defaulted to whatever model is newest, is itself a token-efficiency practice. It is also the same discipline we bring to custom AI agent builds for clients: the budget gets designed before the first token gets spent.