Codex bug on Amazon Bedrock drives up GPT-5.6 Sol costs
A defect in the native Codex CLI prevents explicit prompt caching for the GPT-5.6 Sol model, resulting in cache writes accounting for 85 per cent of estimated spend.
A technical defect in the native Codex CLI has emerged as a cost driver for developers using the GPT-5.6 Sol model via Amazon Bedrock. The issue prevents users from opting into explicit prompt caching, a mechanism designed to reduce costs for agentic coding workloads. According to a GitHub issue filed on 21 August 2026, this limitation has resulted in materially higher costs due to the generation of high volumes of cache-write tokens.
Usage data collected between 5 and 8 August 2026 indicates that cache writes accounted for approximately 85 per cent of the model’s estimated spend. A local Codex session reported 76 Sol requests with 6.709 million cache-write input tokens, while recording zero cached input tokens. This configuration led to an average of about 88,000 cache-write tokens per request. These figures are usage-derived estimates based on Cost Explorer data and the Bedrock rate card, rather than finalised AWS invoice amounts.
The root of the problem lies in the current request types for both HTTP and WebSocket responses, which do not include the necessary controls for the documented explicit-cache mechanism. Although Codex already emits a session-scoped prompt_cache_key, the native Bedrock provider lacks the structured request-body transformation required to utilise it. Consequently, the issue cannot be resolved by editing the standard config.toml file, which only exposes transport and authentication settings.
Amazon documents an explicit cache mode for GPT-5.6 on Bedrock specifically for agentic workflows involving long, stable instructions and tool definitions followed by changing content. This matches the workload described in the report, where the system fails to reuse cached data for stable prefixes. The report notes that while cold starts and genuinely distinct prompts can require cache writes, the current configuration offers no way to leverage the explicit-cache mechanism for the stable-prefix case.
The issue is related to a previous report, number 35300, but adds independent production usage evidence from the native amazon-bedrock provider. The report clarifies that not every cache write is necessarily a defect, as forks and compaction can also require writes. However, the absence of controls for the documented explicit-cache mechanism remains the central concern for users seeking to optimise costs for agentic coding tasks.
