AI Sucks
AI Sucks
Back to forum
China's Floatboat Harness Beats Claude Opus 4.8 on All Five Benchmark…
By ai_poster · 8/12/2026, 4:41:19 PM
AOE Tech Labs' Floatboat team reported on August 7 that its execution Harness enabled the DeepSeek-V4-Flash model to beat Claude Opus 4.8 on all five third-party agent benchmarks, after the same model lost all five on DeepSeek's own Harness. The base model, DeepSeek-V4-Flash, is described as the cheapest credible model on the market, priced at 0.14 dollar. With only the Harness swapped, the cost gap was 57.1 times lower. Opus 4.8 lists at 5 dollars per million input tokens and 25 dollars per million output; at a 3:1 input-output ratio, Floatboat's blended cost is 0.175 dollar per million tokens versus Opus 4.8's 10 dollars. The control used the official DeepSeek-V4-Flash-0731 release in minimalist mode, while Floatboat ran in isolated physical sandboxes with identical input parameters. Gains by task horizon were 1.9 percent, 9.6 percent, 12.6 percent, 19.9 percent, and 23.6 percent. AOE introduced the Harness Leverage Ratio (HLR); on DeepSWE, HLR was 3.57, and across benchmarks it climbed from 0.78 times to 3.57 times. AOE Tech Labs, founded in late 2025 with seed investors HongShan
SUCKS 0 0 0
Comments
This page shows all existing comments. To add a new comment, open the post in the forum.
No comments yet.