OpenAI Resets Codex Quotas After Fixing 8 Token-Draining Bugs
OpenAI has reset usage quotas for paid Codex and ChatGPT Work subscribers after identifying and fixing eight backend issues that caused excessive token consumption.
The changes follow thousands of user reports and a broader audit of token accounting across agentic workflows. Although nominal usage limits remain unchanged, OpenAI estimates that the fixes should allow existing quotas to last 10% to 50% longer, depending on workload characteristics.
The incident also exposes a broader engineering challenge in agentic AI systems: token consumption is no longer determined solely by the user’s visible prompts and model responses. Background memory workers, context compaction, tool retries, subagents, automation triggers, and task summarization can all generate additional model activity.
🔍 Eight Token-Draining Bugs #
OpenAI identified eight separate failure modes that could silently increase token consumption or accelerate quota depletion.
1. Context Compaction Retained Images #
Long-running sessions use context compaction to reduce the amount of historical information carried forward. A bug caused older image assets to remain in the history window after compaction.
Because the retained images continued contributing to context size, sessions could repeatedly cross the compaction threshold. This created a feedback loop in which compaction was triggered more frequently than intended.
OpenAI estimates that image-heavy workflows saw approximately 10% lower token consumption after the fix.
2. Runaway Goals and Infinite Tool Retries #
Some agentic /goal tasks continued executing after their objectives had already been satisfied. Separately, tool failures could cause agents to retry indefinitely rather than terminate or escalate the failure.
The resulting consumption could be substantial. In extreme cases, a single runaway task consumed between 15% and 70% of a user’s weekly quota.
The underlying issue demonstrates why agent termination conditions must be treated as resource-control mechanisms rather than merely workflow logic.
3. Memory Worker Stop-Hook Loops #
Background memory workers inherited Stop hooks intended to terminate processing under specific conditions. When those conditions remained unsatisfied, some workers could repeatedly execute instead of terminating.
The issue affected fewer than 1% of users, but the affected workloads produced severe outliers. One reported task caused a background system check to execute more than 15,000 times.
This class of failure is particularly difficult to detect because the work occurs outside the user’s primary interaction loop.
4. Unauthorized Subagent Model Escalation #
Complex agentic workflows can delegate work to subagents using different models or execution modes. OpenAI found cases where smaller routing models, including Luna, could escalate delegated work to more expensive flagship models without an explicit user request.
There were also cases where a primary model operating in a standard speed mode caused subagents to use /fast mode.
The updated implementation constrains subagents to the model hierarchy and speed tier explicitly selected by the user, reducing unexpected usage escalation.
5. Automation Tasks Triggered Too Frequently #
User-configured automated tasks were sometimes awakened and executed more frequently than their configured schedules specified.
Unlike interactive requests, these executions require no additional user action. As a result, an overly aggressive trigger frequency could accumulate token consumption in the background and deplete quotas without an obvious corresponding increase in visible activity.
The fix aligns background execution frequency with the configured automation schedule.
6. Redundant Computer History Summarization #
Codex can summarize desktop activity to maintain context about computer interactions. Earlier implementations repeatedly summarized highly overlapping portions of historical activity.
For active desktop users, redundant summarization could account for as much as 20% of total weekly usage.
The optimization reduces unnecessary reprocessing by avoiding repeated summarization of information that has already been incorporated into the agent’s working context.
7. Background Rolling Task Summaries #
Normal conversational turns could also trigger additional background requests to generate rolling task summaries.
This behavior introduced approximately 1% additional token overhead across conversational turns. OpenAI has disabled the feature entirely to eliminate the unnecessary background consumption.
8. MCP Double-Encoding #
Results returned from Model Context Protocol (MCP) tool calls could occasionally be encoded twice before being inserted into the model context.
Double encoding increases the effective size of tool output and therefore increases the number of tokens required to process subsequent requests.
OpenAI also identified cases where truncated tool descriptions caused the system to repeatedly retrieve tool schemas, creating another source of redundant context generation.
Fixing both behaviors reduces unnecessary token expansion in MCP-heavy workflows.
🏗️ Architectural Safeguards and Usage Visibility #
The incident illustrates why token accounting becomes substantially more complex in agentic systems.
A single user request can initiate a chain of operations involving the primary model, subagents, external tools, memory workers, context compaction, automation infrastructure, and background summarization. A failure in any one component can therefore create significant hidden resource consumption.
OpenAI is introducing several safeguards to reduce the likelihood and impact of similar regressions.
Architectural Changes #
Low-level changes have been made to the agent scheduling and memory-processing pipelines to prevent recurring execution loops and improve termination behavior.
The goal is to make resource consumption bounded by explicit workflow constraints rather than relying solely on individual components to terminate correctly.
Automated Anomaly Detection #
OpenAI plans to automatically alert engineering teams when token-consumption metrics exhibit abnormal background spikes.
This provides an additional layer of protection against runaway workers, unexpected retry loops, excessive automation triggers, and similar regressions.
In-App Token Usage Analytics #
OpenAI is also developing an in-app usage analytics dashboard intended to provide more granular visibility into token allocation.
The planned breakdown will distinguish consumption across areas such as:
- Primary task execution
- Memory processing
- Subagent workloads
- Background operations
- Other agentic infrastructure
Such visibility would make it easier for developers and power users to determine whether unexpected quota consumption originates from their own prompts or from background agent activity.
📊 What the Quota Reset Means for Developers #
The quota reset does not increase the published nominal limits for paid Codex or ChatGPT Work plans. Instead, the effective capacity of those limits should improve because previously wasted token consumption has been removed.
The largest benefits are likely to appear in workloads that rely heavily on:
- Long-running agent sessions
- Image-rich context
- MCP integrations
- Desktop or computer-use workflows
- Background memory processing
- Automated tasks
- Multi-agent or subagent delegation
- Tool-intensive development workflows
For conventional interactive usage, the improvement may be smaller because several of the bugs disproportionately affected background or highly agentic workloads.
More importantly, the fixes reinforce a fundamental requirement for production AI-agent infrastructure: token accounting must cover the entire execution graph, not just the visible model request and response.
As AI systems become increasingly autonomous, reliable quota management requires the same engineering disciplines used for other distributed systems—bounded retries, explicit termination conditions, resource budgets, anomaly detection, and observability across asynchronous background processes.