Codex on AWS Bedrock Lacks Explicit Cache Controls, Causing High GPT-5.6 Sol Costs

High Costs Due to Missing Cache Controls

The native Amazon Bedrock provider for Codex CLI, version 0.147.0, does not support explicit prompt caching for the openai.gpt-5.6-sol model. This omission causes agentic coding workloads to generate a large volume of cache-write tokens, leading to materially higher operational costs. Observed production usage showed cache writes constituting approximately 85% of the model's estimated spend.

Observed Usage and Impact

For several days in August 2026, an agentic coding workload using GPT-5.6 Sol on Bedrock Mantle Responses API in us-east-1 incurred substantial cache-write costs. A local Codex session reported 76 Sol requests with 6.709 million cache_write_input_tokens and zero cached_input_tokens, averaging about 88,000 cache-write tokens per request. These are usage-derived estimates, not final AWS invoice amounts.

Technical Details of the Issue

While Codex emits a session-scoped prompt_cache_key, the request types for both HTTP and WebSocket Responses requests do not include prompt_cache_options or prompt_cache_breakpoint. The built-in Amazon Bedrock provider configuration only exposes transport/auth settings, preventing the configuration of these caching parameters via config.toml. AWS documentation specifically outlines explicit cache mode for GPT-5.6 on Bedrock for agentic workflows with stable instructions, which matches the affected workload.

Requested Resolution

To address this, support is requested for serializing prompt_cache_options for GPT-5.6-capable Responses providers and adding a typed prompt_cache_breakpoint field to supported input content blocks. Additionally, a provider/model capability gate and a safe placement strategy are needed at the end of Codex's measured stable instruction/tool prefix. Surfacing cache reads and writes in per-turn usage telemetry would also help users diagnose costly full-prefix rewrites.