OpenSquilla tops DRACO with a four-model Chinese ensemble puts the Chinese source story into context for international readers. The point is OpenSquilla argues that routing and consensus across four Chinese models can beat single flagship models on deep-research tasks at lower cost.
What changed
The verifiable facts are: DeepSeek v4, GLM-5.2, Kimi K2.7 and Qwen3.7 in an Agentic Routing flow, Brave Search DRACO score 64.09 versus Opus 4.8 at 59.11 and GPT-5.5 at 53.28, average task cost of $0.12, and DuckDuckGo score 60.85 at $0.39. These details keep the story grounded beyond launch language or market noise.
Why it matters
The harness layer can become valuable when it selects models, collects evidence, merges answers and logs decisions rather than buying only the strongest model. For readers outside China, the signal is also about how AI products are moving from demos into budgets, hardware limits, regulation and operating workflows.
What to watch
DRACO uses LLM judging, so real workflows must confirm long-task failure rates, traceability and enterprise deployment cost. The next useful check is not another headline, but whether the claim holds up in customer deployments, third-party tests or sustained usage.
Sources verified: OpenSquilla GitHub, Agentic Routing technical report and DRACO leaderboard, CocoLoop.