PLATFORMS Signal 505
AWS Bedrock Codex integration reportedly lacks cache controls causing 10x token charges
A bug in the native AWS Bedrock Codex integration prevents explicit cache controls for GPT-5.6 Sol, leading to excessive cache-write token usage and higher costs
Engineers using Codex on AWS Bedrock for agentic coding workloads may face unexpectedly high costs due to uncontrolled cache writes. Without explicit cache controls, stable instruction prefixes trigger repeated cache writes instead of reuse. This issue highlights the need for better provider-level configuration to match documented AWS caching features
Written by elseif from the cluster below · every claim links back to a sourceThe three things worth knowing
Native Codex CLI requests to AWS Bedrock Mantle cannot opt into GPT-5.6 Sol's explicit prompt caching, causing high cache-write token volumes
Cache writes accounted for approximately 85% of estimated model spend in observed production workloads
The issue stems from missing request-body fields for prompt_cache_options and prompt_cache_breakpoint in the native Bedrock provider
THE READ
What the cluster adds up to.
The reported bug affects engineers using Codex with AWS Bedrock's GPT-5.6 Sol model for agentic coding tasks. The core issue is the inability to enable explicit prompt caching, a feature AWS documents for stable instruction prefixes. Without this control, every request with a stable prefix triggers a cache write, multiplying token usage and costs. The material shows cache-write tokens reaching 85% of total spend, a significant deviation from expected behavior where stable prefixes should reuse cached results
The technical gap lies in the native Bedrock provider's request structure. Codex emits a session-scoped prompt_cache_key, but the HTTP and WebSocket request types lack prompt_cache_options and prompt_cache_breakpoint fields. Since the provider config only exposes transport/auth settings, users cannot modify request-body transformations via config.toml. This limitation forces all requests into an implicit caching mode, which is inefficient for workloads with long, stable instruction sets followed by variable content
Adopting a workaround is currently not possible without modifying the provider code. The issue report requests support for serializing prompt_cache_options, adding a typed prompt_cache_breakpoint field, and surfacing cache telemetry in usage metrics. These changes would allow engineers to diagnose and control cache behavior, aligning Codex's behavior with AWS's documented caching mechanisms. Until then, users face a trade-off between cost and functionality for agentic workloads on Bedrock
The cost impact is material, with observed production usage showing cache-write costs of $1,182 out of a total $1,386 for a four-day period. Local sessions mirrored this pattern, with an average of 88K cache-write tokens per request. While cold starts and genuinely distinct prompts will always incur cache writes, the lack of explicit controls prevents optimization for the stable-prefix case. Engineers should monitor their Bedrock usage closely and expect higher-than-anticipated charges until the provider adds caching configuration support
Written by elseif from the cluster below · checked for specifics the sources never containedTHE CLUSTER