The software harness steering an AI coding agent can drive costs up to 40 times higher than the lightest alternative, even when using identical models, according to three recent benchmarking studies published between June and August. The spread stems from a "startup tax" — the system prompt, tool descriptions, and environment setup that each harness ships before any actual work begins — which ranged from around 700 tokens for one harness to roughly 26,000 tokens for another in the June benchmark. Because inference requests require context to be resupplied or reconstructed on every turn, providers process large, overlapping text blocks repeatedly, multiplying that initial overhead across dozens of interactions.
A June benchmark measured twelve harness configurations on the same Python tasks using two unrelated models, reporting token consumption per solved task ranging from approximately 3,500 for one setup to 292,000 for another. The ordering remained nearly unchanged between the two models, pointing to harness software rather than model behavior as the driver. In August, Composio tested eight harnesses with DeepSeek V4 Flash across 30 enterprise workflows spanning Airtable, Gmail, Google Sheets, GitHub, and Slack, finding cost per successful task ranged from $0.028 for Pi Agent to $0.195 for Claude Code. DeepAgents matched Claude Code's pass rate exactly while costing a quarter as much per success. Artificial Analysis tracks harness-model pairings continuously across 326 tasks from three benchmark suites, averaging pass rates over three attempts each and reporting cost per task, token use, and wall time for every pairing.
The June benchmark identified the startup tax as the core mechanism, noting that a harness carrying a 26,000-token floor through fifteen turns spends roughly 390,000 input tokens on scaffolding alone. The report states that startup tax multiplied by turn count predicts tokens per solved task with an R-squared of 0.99 across both models, and concludes that "developers looking to cut agent spend should look first at the prompt floor and the turn count before touching anything more sophisticated." Every harness grew its context at a similar rate of a few hundred tokens per turn, so growth doesn't separate them — the expensive harnesses simply carry a heavier floor. Cache behavior emerged as a second major cost driver: Composio disclosed that Claude Code drew only 1.5% of its input tokens from cache, against roughly 70% for Codex, and since fresh input costs about five times as much as cached input, Claude Code's invoice was far higher than rivals' despite comparable token consumption.
The findings matter because harness choice moves cost by multiples while moving task success by percentage points — different magnitudes rather than different directions. Composio reported pass rates ranging from 46.7% to 66.7% across the eight harnesses on one model, a 20-point spread, while cost spreads reached fourfold or more. The report recommends that procurement teams ask for cost per verified outcome and refuse per-token comparisons, and warns that cache discounts depend on endpoint dialect, gateway, and provider — not the harness alone. Platform teams can verify the cached fraction on their own traffic in an afternoon, and that exercise "is worth more than a model migration," the report notes. For enterprises standardizing an agent platform, the harness now deserves the same scrutiny as the model, since harness choice produced cost spreads large enough to rival major differences in model pricing on workloads where the model held roughly steady. Teams building their own platforms will need to balance prompt efficiency against capability, since the lightest harnesses don't always solve the hardest tasks — and silent truncation under load, observed in some configurations, can look exactly like success in the output.