Codex on AWS bedrock bug causing 10x charges
github.com
August 21, 2026
2 min read
47/100
Summary
A GitHub issue reports that Codex CLI 0.147.0 using its native Amazon Bedrock provider cannot enable explicit prompt caching for the openai.gpt-5.6-sol model through the Bedrock Mantle Responses API in us-east-1. The reporter says Codex emits a session-scoped prompt_cache_key, but its HTTP and WebSocket Responses request types do not support prompt_cache_options or prompt_cache_breakpoint fields. The built-in Bedrock provider configuration supports transport and authentication settings rather than request-body transformations, preventing configuration through config.toml. For completed days from August 5 through August 8, 2026, the reporter estimated 3,656 requests generated 171.94 million cache-write tokens, costing $1,182.09 of an estimated $1,386.46 total model spend. Cache writes represented about 85% of the estimate, which was derived from AWS Cost Explorer quantities and Bedrock rate-card pricing rather than finalized invoice amounts. A local session recorded 76 Sol requests, 6.709 million cache-write input tokens, no cached input tokens, and roughly 88,000 cache-write tokens per request; corresponding CloudWatch metrics showed no client errors. The requested changes include support for prompt_cache_options, typed prompt_cache_breakpoint input fields, provider and model capability checks, placement after stable instruction and tool prefixes, and per-turn cache-read and cache-write telemetry.
Key Takeaways
- Codex CLI's native Amazon Bedrock provider currently lacks request fields needed to enable GPT-5.6 Sol explicit prompt caching, according to GitHub issue #37674.
- The issue reporter estimated that 171.94 million cache-write tokens accounted for about $1,182.09, or 85%, of $1,386.46 in Sol spending over four completed days in August 2026.
- A local Codex session reported zero cached input tokens alongside 6.709 million cache-write input tokens across 76 Sol requests.
- The requested implementation would expose explicit cache controls and per-turn cache-read and cache-write usage telemetry.
What the discussion said
The thread centered less on the headline alone than on a costly failure in Codex’s Bedrock prompt caching: users saw cache reads below 5%, expensive writes that did not pay off, and bills reportedly approaching ten times the expected level. Several readers treated disabling web search as a practical escape hatch, but they were irritated that such a consequential workaround was buried rather than clearly surfaced. Others said Codex more broadly has been consuming usage unusually fast, reinforcing the sense that metering and caching behavior have become opaque at exactly the point customers need trustworthy cost controls. The technical discussion was muddled by an issue thread some readers found incoherent and apparently AI-authored. One reader nevertheless found the apparent change in cache behavior architecturally interesting, speculating that a newer model design may no longer support conventional prefix-state reuse. That possibility sharpened the criticism: if a major inference or caching change altered customer economics, the provider should explain it plainly. The sharpest split was over what these operational mistakes say about AI’s readiness to replace software developers. Skeptics saw a basic chat-product billing bug as evidence against grand automation claims; the opposing view was that poorly executing incumbents, not AI itself, are the vulnerable part of the equation.
Where opinion split
Do repeated Codex billing and caching failures undermine claims that AI can replace software development? Skeptics argue that a company selling frontier AI still mishandling straightforward product behavior exposes how premature those claims are. The counterargument is that such failures show incumbent companies are ripe for AI-driven disruption, rather than demonstrating a limit of the technology.
Community Sentiment
Positives
- Disabling Codex web search reportedly restores sane Bedrock cache behavior, giving affected teams an immediate way to stop runaway inference costs.
- The odd cache behavior sparked useful architectural scrutiny: a model-state redesign could explain why familiar prefix caching no longer works as users expect.
Concerns
- Bedrock users reported cache reads under 5% alongside costly unused writes, turning what should be efficient Codex prompting into bills roughly ten times higher.
- The discussion around the bug was criticized as confusing and insufficiently explained, leaving customers unable to tell whether caching mode, breakpoints, or model behavior caused the charges.
- Reports of Codex rapidly burning through usage suggest the billing problem may not be isolated, making AI coding-assistant costs feel unpredictable.
- Readers saw the failure of a frontier-AI provider to manage basic chat-product economics as strong evidence against claims that AI is ready to replace software engineers wholesale.