HarnessTax: How Much Does the Harness Matter for Coding Agents?

The same model often achieves a similar success rate at substantially different costs. GPT-5.6 Luna offers the lowest cost on both benchmarks, while Claude Fable 5 reaches the highest success rate on SWE-bench Lite. Kimi K3, an open-weight model, is close to the Pareto frontier near GPT 5.6 Sol in SWE-Bench Lite, and sits just below the Pareto frontier on Terminal-Bench 2.0. However, these models show no substantial performance differences across harnesses. Claude Fable 5 solves 97.8% of attempts in Claude Code, 96.7% in Codex and 96.7% in Pi, yet Claude Code costs about twice as much as Pi ($1.33 vs $0.67).

Fable 5 achieves a slightly higher success rate in Claude Code than in Pi, at about twice the cost. The cost gap extends beyond Fable 5. Across shared models, Claude Code costs about 2.0× as much as Pi and 1.6× as much as Codex on SWE-bench Lite, and 1.5× as much as Pi on Terminal-Bench 2.0, using geometric means of cost ratios. Meanwhile, the average harness effect on success rate stays within ±2% on SWE-bench Lite and within about ±5% on Terminal-Bench 2.0.

Paying extra for essentially the same quality because the use of different harnesses is like paying a… Harness Tax 💰...[10] And you may be paying such a hidden “harness tax” when you accept a coding agent’s default harness without comparing alternatives. Model evaluations should therefore compare the same model’s cost and task success across commonly used harnesses.

Pi reaches the Pareto frontier on both benchmarks by providing just four tools: read, write, edit, and bash [11]. To understand how harness design affects spending, we examine costs across completed attempts, recorded turn counts, and initial context.

Agents can take similar numbers of turns at substantially different costs. For Fable 5 on SWE-bench Lite, Pi and Claude Code average 15.4 and 15.3 turns per attempt, yet Claude Code costs about twice as much for a 1.1% increase in the success rate. This means higher spending per recorded turn, though turn definitions vary across harnesses.

A harness tax can begin with the first model call. Across all seven models, Claude Code’s mean initial context is over 10× Pi’s, with longer instructions and larger tool schemas. This additional context can raise costs, though total spending also depends on caching, generated tokens, and later calls.

The effectiveness of Pi and Codex demonstrates opportunities for open-source harness research with existing models. Researchers can work with SOTA coding harnesses without access to proprietary harnesses or co-training with the model. Richer harness features may still benefit other models, workloads, or interaction settings. Harness complexity should therefore be treated as an empirical trade-off.

Provider-specific optimization does not guarantee the best pairing. Providers sometimes optimize models for their coding environments: OpenAI, for example, describes GPT-5-Codex as optimized for software engineering in Codex [12]. Yet across the six Anthropic and OpenAI models and both benchmarks, an alternative harness achieves the highest observed success rate in nine of twelve comparisons.

The result extends beyond Opus and beyond Claude Code. Sonnet 4.6 solves 68.9% of attempts in Codex versus 66.7% in Claude Code on SWE-bench Lite at a similar cost. GPT-5.6 Sol also performs competitively outside its provider's harness: on Terminal-Bench 2.0, it achieves an 83.3% success rate in Pi versus 78.9% in Codex, at about half the cost ($0.42 versus $0.76). Across the six Anthropic and OpenAI models and both benchmarks, an alternative harness achieves the highest observed success rate in nine of twelve comparisons.

These results show that a model’s capabilities are compatible, generalizable and can carry over to other harnesses. Providers do report to optimize some models for their own coding environments: OpenAI, for example, describes GPT-5-Codex as optimized for agentic software engineering in Codex [12]. Yet we observe that a shared provider does not guarantee the best pairing. The practical question remains which harness delivers the best balance of cost and task success for a given model and workload.

Our results show that the same model can achieve similar success rates at substantially different costs. On the benchmarks we test, simple open-source harnesses can be competitive, and models can perform well outside their own harness. A harness tax can go unnoticed when we focus only on task success.

Across models, Pi and Codex often achieve similar success at lower cost than Claude Code. These findings can be limited to the two open-source benchmarks we test, which the models may have encountered during training. Results may differ on other benchmarks and workloads.

Much prior work, including our work on retrieval agents [13], shows the value of choosing models and system configurations together. Harness selection is an even more pressing problem for coding agents given the volume and spread of its usage. The natural next step is to evaluate harnesses and automate their selection in real development workflows, where requirements evolve, developers provide feedback [14], and tasks extend across sessions [3].

More broadly, the need for different harnesses depends on the role of coding agents. For day-to-day tasks, coding agents are essentially interfaces to model intelligence: they manage context, access tools, and execute tasks [15]. As models become more capable, coding agents may need less of today’s scaffolding. General-purpose coding agents should therefore prioritize cost efficiency and reliability, as many tasks may not require fancy add-on features. For harder problems at the boundary of a model’s capabilities, including scientific discovery, (coding) agents may still benefit from harnesses that provide structured guidance for exploring ideas, evaluating candidates, and learning from feedback. Harness research can be viewed as a way to help models push the boundaries of knowledge, unlocking the next phase of intelligence in the process. Yet users should not have to make these configuration decisions themselves. We should envision a redesigned harness that adapts as tasks unfold while remaining general.

If HarnessTax is useful in your research or work, please cite the project as: