OpenSquilla 集成方案登顶 DRACO Brave 组，成本仅为 GPT-5.6-sol 十四分之一
OpenSquilla 用四款国产模型协同，以 1/14 成本达到超越 GPT-5.6-sol 的质量分，验证多模型集成路线。
DRACO 评测 Brave 组中，OpenSquilla 集成方案以 64.09 分居首，成本 $0.12/任务；GPT-5.6-sol 得分 63.99，成本 $1.71。方案由四款国产模型并行提案、一模型聚合，未用海外旗舰，显示复杂 Agent 任务正从拼单模型转向拼组织。
正文摘录
Recently, in the Brave Search group of the public DRACO benchmark, the overseas flagship model GPT-5.6-sol entered the leaderboard with an average score of 63.99 and an average task cost of $1.71. OpenSquilla 0.5.0 Preview's multi-model ensemble scored 64.09 on average, still ranking first in the group, with an average task cost of just $0.12. In other words, while its quality score is essentially on par with GPT-5.6-sol — 0.10 points higher, to be precise — OpenSquilla's average task cost is about 1/14 of GPT-5.6-sol's, and roughly 1/10 of Fable 5's $1.21. The ensemble uses four domestic models (DeepSeek, GLM, Kimi, Qwen) to propose answers in parallel, then a single model aggregates the final output. No overseas flagship is in the lineup. As of now, the Brave Search group comparison has …