Native Bedrock Codex GPT-5.6 Sol lacks explicit cache controls, producing high cache-write spend · Issue #37674 · openai/codex · GitHub

Summary

Native Codex CLI requests to Amazon Bedrock Mantle cannot opt into GPT-5.6 Sol explicit prompt caching. On an agentic coding workload, this has produced a large volume of cache-write tokens and materially higher cost.

This is related to #35300, but adds independent production usage evidence from the native amazon-bedrock provider.

Environment

• Codex CLI: 0.147.0
• Provider: native amazon-bedrock
• Endpoint: Bedrock Mantle Responses API, us-east-1
• Model: openai.gpt-5.6-sol

Observed production usage

For the completed days 2026-08-05 through 2026-08-08, Cost Explorer usage quantities and the Bedrock rate card produced the following cache-aware estimate for Sol:

Cache writes were about 85% of the model’s estimated spend.

A local Codex session also reported 76 Sol requests with 6.709M cache_write_input_tokens, zero cached_input_tokens, and an average of about 88K cache-write tokens per request. There were no client errors in the corresponding CloudWatch metrics.

These are usage-derived estimates, not finalized AWS invoice amounts.

Investigation

Codex already emits a session-scoped prompt_cache_key, but the request types for both HTTP and WebSocket Responses requests do not include either:

• prompt_cache_options
• prompt_cache_breakpoint

The built-in Amazon Bedrock provider config exposes transport/auth settings, not structured request-body transformation, so this cannot be configured through config.toml.

AWS documents explicit cache mode for GPT-5.6 on Bedrock specifically for agentic workflows with long stable instructions/tool definitions followed by changing tool and user content. That matches the workload above.

Requested behavior

1. Add support for serializing prompt_cache_options for GPT-5.6-capable Responses providers.
2. Add a typed prompt_cache_breakpoint field to supported input content blocks.
3. Provide a provider/model capability gate and a safe placement strategy at the end of Codex’s measured stable instruction/tool prefix.
4. Surface cache reads and cache writes in per-turn usage telemetry so users can diagnose costly full-prefix rewrites.

Scope

This report does not claim that every cache write is a defect. Cold starts, genuinely distinct prompts, forks, and compaction can all require writes. The issue is that native Bedrock Codex currently has no way to use the documented explicit-cache mechanism for the stable-prefix case.