The Chinese team AOE Tech Labs, which developed the agent desktop product Floatboat, published full Harness evaluation data across five third-party agent benchmarks on August 7. The base model used was DeepSeek-V4-Flash, which is the most affordable reliable model on the market. Only the executing Harness changed while keeping the model and cost constant. The model, which previously showed results lower than Claude Opus 4.8 across all five benchmarks in its own DeepSeek Harness, managed to surpass Opus 4.8 within the Floatboat Harness.
This result is even more significant considering the pricing. Opus 4.8 costs $5 per million input tokens and $25 per million output tokens. At an input-to-output ratio of 3:1, commonly found in agent workflow scenarios, the combined cost of Floatboat is only $0.175 per million tokens, whereas for Opus 4.8 it is $10, representing a gap of 57.1 times. These five benchmarks are public leaderboards from third parties, including BrowseComp created by OpenAI. The data is not an artifact of question selection.
The research methodology was conducted with maximum rigor. A control group used the official version of DeepSeek-V4-Flash-0731, not a preliminary checkpoint. The control group operated in the provider's own Harness in a minimalistic mode, making the comparison stricter as the provider's engineers were involved in the process. The Floatboat side operated in isolated physical sandboxes with identical input parameters. The only variable was the Harness itself.
When sorting tasks by execution horizon from short to long, the performance gain does not decrease: 1.9 percent, 9.6 percent, 12.6 percent, 19.9 percent, and 23.6 percent. Short tasks are essentially single-prompt question-and-answer sessions where model quality plays a decisive role. Long-horizon tasks, which reflect real work, depend on the execution system: file access, tool calls, state saving, multi-step self-correction, and rollback. Real work occurs at the end of the long horizon.
AOE Tech Labs introduced the Harness Leverage Ratio, or HLR, to make the comparison verifiable. The calculation formula is the gain achieved solely by changing the Harness divided by the gain achieved when upgrading to a stronger reference model. An HLR value greater than one means that the Harness itself surpasses the model upgrade. On DeepSWE, Floatboat reported an HLR of 3.57: by keeping the model unchanged and only changing the Harness, a 3.57 times greater public performance was achieved than by upgrading to Opus 4.8. Across all five benchmarks, the HLR monotonically increased with the increase in horizon length from 0.78 to 3.57.
AOE Tech Labs was founded at the end of 2025 with participation from investors HSG and VLight Capital in a seed funding round. The team released Floatboat Desktop in January, FloatIM in April, and FloatSchedule in May. The core idea is that a system that continuously narrows the gap between the evolving target execution system and the drifting model output is key to everyday agents. The cheapest model can now match or exceed the most expensive cutting-edge model when the Harness is designed for long-horizon operation.