← Back to feed

OpenAI’s GPT-5.6 upgrade to Codex has produced an unexpected billing crisis for power users. Subscribers on the $200/month Pro plan — which includes a 20x Codex allowance — are reporting their 5-hour weekly allowance disappears after a handful of tasks that would previously have consumed a fraction of that budget under GPT-5.5.

The pattern is consistent across multiple user reports on the OpenAI Developer Community forum. Before the GPT-5.6 update, Pro subscribers regularly ran GPT-5.5 with Extra High reasoning through long development sessions over multiple days without approaching their quota. After the update, comparable task types drain the weekly allowance within hours. One detailed account documents approximately 3,500 purchased credits — over $100 in paid charges above the monthly subscription — consumed in a single session while the user’s weekly allowance was still largely intact.

What’s Being Reported

The anomaly isn’t just that GPT-5.6 uses more tokens — it’s the scale of the change and the way it interacts with quota accounting. Users describe the weekly 5-hour allowance disappearing after “a couple of tasks,” and purchased credits being drawn down in parallel with the included allowance rather than as a fallback. One user on the forum noted that banked rate-limit reset capacity was also available and unused, yet purchased credits were consumed first.

The specific mechanism is undocumented by OpenAI. GPT-5.6’s architecture — particularly its Sol, Terra, and Luna variants with parallel subagent capabilities — generates significantly more intermediate tokens per completed task than GPT-5.5. Each Codex session invokes planning, tool calls, verification, and rollback loops. If GPT-5.6 issues more verification steps or runs parallel traces by default, the token multiplication effect compounds quickly.

OpenAI’s Response

Support responses cited in the forum thread confirm the usage as “normal” without addressing why equivalent workloads now cost significantly more. No public statement, pricing clarification, or system card update has followed the Codex GPT-5.6 transition. The OpenAI Developer Forum thread is accumulating confirmations from multiple $200/month subscribers.

Commercial Stakes

The disruption matters because Codex Pro’s flat-rate model was its primary selling point for heavy users — predictable monthly cost, generous compute headroom. If GPT-5.6’s default behaviour effectively voids that predictability, the value proposition shifts. $200/month for a subscription that now triggers $100+ in overage charges per session is structurally different from $200/month for unlimited agentic sessions.

Key Numbers

• Monthly subscription: $200 (Pro plan, 20x Codex allowance)
• Reported overage: $100+ in purchased credits on a single session
• Weekly allowance fate: Drained in hours vs. days under GPT-5.5
• OpenAI response: “normal” usage per support

GPT-5.6’s Sol variant — which leads Terminal-Bench 2.1 at 78.2% and introduced parallel subagent execution — is the most likely culprit. It was not designed as a drop-in replacement for GPT-5.5 in terms of per-task token economics. Whether OpenAI will retroactively document the consumption difference or adjust Pro plan limits remains unaddressed.