OpenAI Found Eight Ways Codex Was Burning Paid Usage

OpenAI says it found eight ways Codex and ChatGPT Work could consume paid usage without delivering useful work, from repeated compaction and background workers refusing to stop to tool results being encoded twice. It reset usage for all paid users and says the fixes should make allowances stretch 10% to 50% further. That estimate is not independently verified; the bigger story is a meter sometimes charging for machinery spinning behind the wall.

OpenAI Watch analysis

What happened

• Compaction retained old images and could trigger compaction again; OpenAI reports roughly 10% lower usage for heavy image users after the fix.
• Memory workers inherited Stop hooks and sometimes refused to finish. One thread reportedly checked whether it could stop 15,000 times.
• Completed goals and broken tools could keep running or retrying, consuming 15% to 70% of a weekly allowance in examples OpenAI examined.
• Some custom automations ran more frequently than configured.
• Smaller subagents could silently choose more capable helpers or fast-mode execution.
• Computer History repeatedly summarised overlapping activity, costing some users up to one fifth of a weekly allowance, according to OpenAI.
• Ordinary turns triggered extra rolling summaries, adding about 1% to usage. OpenAI disabled them.
• MCP results could be encoded twice, while truncated tool instructions were fetched again.

Why it matters

Task cost already varies with repository size, context, tools and model choice. Hidden retries and runaway background work are different: customers cannot manage behaviour they cannot see, or know whether a weekly limit funded their task or an immortal housekeeping process.

The reset is the right immediate remedy and confirms rapid allowance drain was not always the user's fault. OpenAI says it added regression controls and paging, and is working on in-app visibility showing where usage goes.

Our read

The candid bug list is useful: it names failure modes, gives impact ranges and separates shipped fixes from future visibility work. That beats quietly widening allowances and pretending the meter had always worn its trousers correctly.

The claimed improvement remains supplier-reported. The durable fix is itemised accounting for model work, compaction, memory, subagents, automations and retries, so customers can challenge obvious waste.

What to watch

• Whether users see the promised gains across image-heavy, automated and subagent-heavy workflows.
• Whether OpenAI ships per-task and background-worker usage visibility, not another aggregate dial.
• Whether future runaway loops automatically restore affected allowances.
• Whether accounting works cleanly across Codex, ChatGPT Work and shared pools.

Discussion spark: Which usage category should OpenAI expose first: compaction, background memory, subagents, automations or tool retries?

OpenAI Watch is independently operated by WittyWires. It is not affiliated with, endorsed by, or operated by OpenAI.