Urgent.News

What's breaking now, across thousands of outlets.

Tech

Codex on AWS bedrock bug causing 10x charges

An issue with Amazon's AWS Bedrock service is causing increased costs for certain AI coding workloads, according to recent findings. The Native Codex CLI does not allow for GPT-5.6 Sol model optimization through explicit prompt caching. This results in a large number of cache-write tokens, leading to significantly higher expenses.

The problem was observed across a short period from August 5 to August 8, 2026, with 3,656 requests and 171.94 million cache-write tokens. This equates to an estimated cost of $1,182.09, with the total cost amounting to $1,386.46. Approximately 85% of the model's estimated spend was attributed to cache writes.

The native `amazon-bedrock` provider, when used with Codex CLI version `0.147.0`, is affected. The endpoint used was Bedrock Mantle Responses API, located in the `us-east-1` region, with the model `openai.gpt-5.6-sol`.

AWS's own documentation recommends using explicit cache mode for GPT-5.6-capable models in agentic workflows involving long, stable instructions or tool definitions. This matches the described workload. The issue lies in the lack of support for serializing `prompt_cache_options` and adding a `prompt_cache_breakpoint` field. The AI service cannot be configured through `config.toml` due to transport and authentication settings not being exposed.

While not all cache writes are considered defects, the current native Bedrock Codex setup lacks the ability to utilize the documented explicit-cache mechanism for stable prefixes. The proposed solution includes adding support for `prompt_cache_options` serialization, introducing a `prompt_cache_breakpoint` field, and implementing a capability gate and safe placement strategy at the end of the stable instruction/tool prefix.

Additionally, surface cache reads and writes in per-turn usage telemetry to help users identify costly full-prefix rewrites.

Written by urgent.news from Hacker News's reporting — not their text. Machine-written — may contain errors; check the original before relying on it.

Read the original at github.com →

More in Tech

More from Friday 21 August →