Codex CLI 0.147.0 版本在 AWS Bedrock 上使用 GPT-5.6 Sol 模型时存在缓存控制缺陷，造成大量不必要的成本支出1。该工具无法启用显式提示缓存功能，在代理编程工作负载中导致成本增加约 10 倍1。
根据 2026 年 8 月 5 日至 8 月 8 日的测试数据，76 个 Sol 请求产生了 6.709 百万个缓存写入令牌，但缓存读取令牌为零1。缓存写入成本占总支出的约 85%，平均每个请求约消耗 88 千个缓存写入令牌1。问题根源在于 HTTP 和 WebSocket 请求类型中缺少 prompt_cache_options 和 prompt_cache_breakpoint 字段1，导致系统无法正确管理缓存策略。
Codex CLI version 0.147.0 contains a caching control defect when using the GPT-5.6 Sol model on AWS Bedrock that prevents the explicit prompt cache feature from being enabled, resulting in dramatically inflated costs.1 During a test period from August 5 to August 8, 2026, 76 Sol requests generated 6.709 million cache write tokens with zero cache read tokens, averaging approximately 88,000 cache write tokens per request.1 Cache write operations account for roughly 85 percent of estimated spending, driving total costs up by approximately tenfold compared to expected usage.1
The root cause stems from missing prompt_cache_options and prompt_cache_breakpoint fields in both HTTP and WebSocket request types.1 This absence of the necessary cache control parameters prevents the system from properly managing cache utilization, leading to excessive write operations that generate substantial unnecessary charges.
评论
还没有评论，欢迎留下第一条。