记录当前网关目录、模型能力、Token 成本与运行表现。没有执行的测试会明确标为"未实测"。 Catalog, model capability, token cost and runtime records. Tests not yet run are explicitly marked "Not tested".
事实抽取、指令遵循、多轮一致性、中文惯用语与结构化摘要。Extraction, instruction following, multi-turn consistency, idioms and structured summarization.
32k / 64k / 128k 固定文档长度下的压缩率与关键事实保留率。Compression and fact retention at 32k / 64k / 128k fixed document lengths.
bug 修复、单元测试生成与结构化补丁，Python / TypeScript 双语言。Bug fixes, unit-test generation and structured patches in Python and TypeScript.
JSON Schema 合规率与重试次数，覆盖 5 类常用 Schema。JSON Schema compliance and retries across 5 common schema types.
首字延迟、总响应时间与成功率，24 小时周期重复采样。First-token latency, total time and success rate with 24h repeated sampling.
输入/输出价格、典型任务消耗与月度成本测算框架。Input/output prices, typical workload cost and monthly cost framework.
当前网关模型目录的统一记录表，各维度指标待实测后填入。Unified record table for the current gateway catalog; metrics filled after testing.
输入、输出、缓存输入、上下文与官方来源，无法核验时显示待确认。Input, output, cached input, context and official sources — "To verify" where unmatched.
主模型超时或限流时自动切换备选模型，提高可用性。Switch to a backup model on timeout or rate limit to improve availability.
需要核对你的实际模型 ID？Need to verify your actual model ID? 联系支持Contact Support
