Codex Bug on AWS Bedrock Causing 10x Higher Charges
Lack of cache controls in GPT-5.6 Sol model leads to unexpected cost spikes.
1Cost-Increasing Bug in AWS Bedrock
A critical flaw was discovered in Amazon's cloud AI platform AWS Bedrock, where Codex CLI lacks cache control mechanisms when working with GPT-5.6 Sol model. This issue particularly affects workloads with long, stable instructions, causing unexpected high costs for users. Analysis between August 5-8, 2026 showed cache write operations accounted for 85% of total expenses.
2Technical Details and Impact
The bug occurs because while Codex generates session-scoped prompt_cache_key, the AWS Bedrock provider cannot utilize this key. A local Codex session recorded 6.7 million cache write tokens across 76 Sol requests, with zero cache reads. Average 88K tokens written per request significantly inflated costs unexpectedly for users.
3Inconsistency with AWS Documentation
AWS documentation states GPT-5.6 on Bedrock supports explicit cache mode especially for workloads with long stable instructions. However, current native Bedrock Codex implementation cannot use this feature. This reveals inconsistency between AWS's documented capabilities and actual implementation.
4Cost Analysis and User Impact
Cost Explorer data showed cache writes dominated expenses between August 5-8, 2026. Users faced charges up to 10 times higher than expected. Large-scale AI workloads run by enterprises could suffer significant financial losses due to this issue.
5Root Cause Analysis
The problem stems from AWS Bedrock provider configuration exposing only transport and auth settings without request body transformations. Users cannot configure necessary cache settings via config.toml. This technical limitation particularly hinders cost optimization for agentic workloads.
6Comparison with Other AWS Services
AWS typically provides comprehensive cost management tools for similar AI services. However, this Bedrock bug makes cost predictions especially challenging for the new GPT-5.6 model. Users cannot fully anticipate actual costs until billing period ends.
7Temporary Workarounds
Experts recommend affected users to break workloads into smaller chunks and use shorter sessions. Regular monitoring of CloudWatch metrics is essential to detect abnormal token usage. An urgent patch from AWS is expected to address this issue.
8Importance of AI Cost Management
This incident highlights the critical importance of cost management in cloud-based AI services. Users must actively utilize cost monitoring tools, especially with newly released models, and prepare for unexpected charge spikes.