Harness Software, Not Just Models, Drives 70-Fold Variation in AI Coding Agent Token Costs

When evaluating the price of an AI coding agent, most attention flows toward the model itself. Yet recent testing reveals a potentially larger cost driver: the harness—the orchestration layer that guides the model through a task.

Three independent studies compared how different harnesses handle identical coding challenges on the same models, uncovering efficiency gaps that dwarf typical model price differences. The inconsistency points not to model behavior but to how each harness structures its prompts, manages context, and shepherds multi-turn workflows.

The Context Overhead Problem#

Every time a coding agent takes a step, it must resubmit or regenerate context for the next move. This includes the system prompt, tool definitions, and environment setup—what researchers call the "startup tax."

A June benchmark of twelve harnesses on the same Python tasks, all running through the same API on identical models, found startup overheads ranging from roughly 700 tokens (Aider in architect mode) to 26,000 tokens (OpenClaw). While a single 40x cost multiplier might be manageable, the math breaks down in multi-turn loops: a 26,000-token floor compounded across fifteen turns wastes roughly 390,000 tokens on scaffolding alone.

The pattern held steady across two unrelated models—DeepSeek V4 Flash and Nvidia's Nemotron 3 Ultra—suggesting harness design, not model quirks, explains the gap. Researchers found that startup overhead and turn count together predict total token consumption with 99% accuracy. For teams trying to trim agent expenses, lowering the prompt floor and reducing iteration count matter far more than swapping models.

Cache Strategy Changes the Invoice#

Token counts alone mask a second lever: how well each harness exploits cached context. Modern LLM APIs charge a fraction of the normal rate for repeated ("cached") tokens, yet harnesses use this feature wildly differently.

In Composio's August benchmark across eight harnesses on enterprise workflows, Claude Code generated only 1.5% cache hits, while Codex achieved 70% and another harness reached 57%. Since cached input costs roughly one-fifth as much as fresh input, Claude Code's raw token count—though lower than competitors—translated to a higher bill per successful task.

The gap stems not from Claude Code's prompts but from the serving path: Claude Code used Anthropic's message dialect through an API gateway, which apparently reduced cache eligibility compared to an OpenAI-style endpoint. This means cache efficiency depends on routing and provider configuration, not just the harness software itself.