Cutting AI Cost in OpenClaw: The Context Engine, Model Mixing, and Governance

Where the money goes

OpenClaw is an open-source agent that runs real work across many steps, reaching for tools, notes, and files as it goes. The model it runs on has no memory between steps, so on every single step a part of OpenClaw called the context engine assembles the exact package the model will see. Whatever it packs, you pay for. And because it re-packs on every step, anything left in the package is paid for again and again. That re-packing decision is where most of your cost is made, and it is where most of your savings live.

The bill itself comes from text. Models read and write in small chunks called tokens (roughly three-quarters of a word each). You pay for tokens sent in (the package the context engine assembles) and tokens written back (the model's reply), and at most providers the reply costs several times more per token than the input. So three forces drive cost: how heavy the standing package is on every step, how much old material rides along, and how much the model writes back. This report walks through forty concrete levers across those forces, then the model-choice decisions and governance that hold the savings in place over time. The single most powerful idea is also the simplest: tune what the context engine includes, because that is the one cost surface every other cost flows through.

1. What you are paying for

OpenClaw runs on AI models, and you pay for text in both directions. The more text flows, the more you pay. Knowing which text is expensive tells you where to aim.

Every request has two halves. The input is everything the context engine packs and sends to the model: standing instructions, recent conversation, relevant saved notes, the files in play, and the results of tools. The output is what the model writes back. Input tokens are cheaper per token; output tokens cost several times more at most providers. A third, quieter cost is reasoning tokens, the hidden thinking some models do before they answer, which some providers bill as output and others fold into overhead.

Two patterns make cost climb faster than people expect. First, the standing package is paid on every step, so a heavy package multiplies across a long session. Second, history grows: without intervention a long conversation re-sends more and more of itself each turn, so a thirty-turn session can consume over a million tokens in total even though no single turn looked alarming. Most of this report is about flattening those two curves while keeping the work trustworthy.

2. The context engine in depth

This is the heart of the bill. On every step the context engine assembles what the model sees. Tuning that assembly is the single highest-leverage thing you can do, and it starts with deciding how often the standing context is re-injected.

On each step the engine reaches for the standing instructions, the recent conversation, relevant saved notes, the files in play, and tool results, then packs them into one request. Because it re-packs every step, the standing material is the part to watch: it is paid repeatedly. The first knob is whether OpenClaw re-injects your workspace bootstrap files (the standing instructions, described in the next section) on every turn or more selectively.

2.1 Context injection mode

OpenClaw exposes a configuration key, contextInjection, that controls when bootstrap files are re-added to the system prompt. Three modes exist. "always" (the default) injects all bootstrap files on every single turn, so every follow-up message in a long session pays the full standing tax again. "continuation-skip" injects on the first turn and after any user prompt that follows a completed model response, but skips simple back-to-back continuations, removing duplicate injection. "never" performs no automatic injection at all; the agent must supply its own context, which is a large saving but only safe for specialized stateless agents.

A community case study measured roughly a 25 percent reduction in prompt tokens on multi-turn chats from this one change, and one report cited it among a handful of changes that took a monthly bill from over $1,500 to under $50. Volatile The exact percentage depends on session length and file sizes, so treat those numbers as direction, not a promise.

2.2 Startup context

Separately, OpenClaw can inject a one-time prelude at the very first turn of a session, the startupContext, which can pull in recent daily memory logs (files named by date). The default is the last 2 days, up to 1,200 characters per file and 2,800 characters total. This buys session continuity but costs tokens on every session start, so for agents whose task is always fully specified in the request it is pure overhead.