Token metering loop causes physically impossible quota consumption (free/plus/pro)

What version of the Codex App are you using (From “About Codex” dialog)?

26.616.3309.0

What subscription do you have?

varies

What platform is your computer?

_No response_

What issue are you seeing?

Description
A severe backend token‑accounting bug is causing user quotas to be exhausted instantly—often in 1–2 messages—by reporting token counts that violate the physical limits of model inference speed and context‑window size. This affects free, Plus, and Pro accounts, both new and old.

I did the math from the numbers you gave me.

On the new account, 1.9M + 3.0M = 4.9M, which is 100,000 tokens above the 4.8M lifetime display. That alone is a mismatch, though it could be partly rounding if the lifetime number is only approximate.

The bigger problem is the implied rate. If 3,000,000 tokens were burned in 175 seconds (2m 55s), that is about 17,143 tokens/second. On the old account, 860.5M tokens over 9,780 seconds (2h 43m) is about 87,986 tokens/second. If the 7.2B figure was also over that same kind of task window, it would imply about 736,196 tokens/second. Those rates do not look like normal literal chat-token usage.

The week-by-week pattern also jumps hard: 392M → 1.2B = 3.06×, then 1.2B → 1.6B = 1.33×, then 1.6B → 5B = 3.13×, then 5B → 7.2B = 1.44×. That looks more like a quota counter, cumulative meter, or corrupted usage display than a clean per-message token count.

So the deduction from your stats is: either the thing you are looking at is not literal text tokens, or the usage counter is broken/misreported. Your numbers do not support the idea that a couple of short chat messages literally consumed billions of normal chat tokens.

What steps can reproduce the bug?

Use “extra high” reasoning mode in a fresh conversation (no plugins, no sub‑agents).

Send a single, moderate‑length message (e.g., 2–3 paragraphs).

Allow the model to respond; do not cancel the task.

Check the token usage counter immediately after the reply.

Observed Behavior

A single incomplete message reports 3,000,000 tokens consumed in 175 seconds (longest task = 2m55s).

Another message on a different day reports 1,900,000 tokens in ~1.5 messages.

Token counter jumps from 0 to 4.8M “lifetime” across only 5 total threads and no plugin/skill usage.

On an older account, the counter repeatedly displays 7,200,000,000 tokens drained in “a few messages”, and even “compensation” quotas instantly hit the same 7.2B figure.

The “peak token” on a free account was 860,500,000 in a 2h43m task.

What is the expected behavior?

Token consumption should correspond to the actual number of tokens in the prompt and the generated response.

For a single‑threaded chat without plugins, maximum consumption per message cannot exceed prompt_length + response_length, bounded by the model’s context window (e.g., 128k tokens).

Generation rate cannot exceed ~100–200 tokens per second for a single stream.

Evidence of Physical Impossibility

Output speed exceeds hardware limits

3,000,000 tokens / 175 s = 17,143 tok/s

860,500,000 tokens / 9,780 s = 88,000 tok/s
Publicly known LLM inference for GPT‑5.4‑class models maxes out at ~200 tok/s per stream. Even with 50+ parallel sub‑agents, 88k tok/s is unachievable.

Per‑message token count exceeds context window

1,900,000 tokens in 1.5 messages ≈ 1,270,000 tokens per message
The maximum prompt+response size for GPT‑4 is 128k tokens; 1.27M is ~10× that ceiling.

Internal counter contradictions

“Lifetime” shows 4,800,000, but Day 1 + Day 2 usage = 4,900,000 (overshoots the displayed total).

A single 2‑minute task represents 62.5% of all lifetime tokens.

The number 7,200,000,000 appears identically across multiple “resets” and compensation quotas, indicating a stuck accumulator.

Usage progression breaks physics

Old account weekly totals jump from a plausible 1.6B (with many sub‑agents) to an instantaneous 7.2B drain, with no gradual increase.

Impact
All usage becomes impossible within 1–2 messages. Users cannot use the product at all. This is being reported widely by free, Plus, and Pro users (see related issues: [link to the “x20 their message drain” GitHub thread you mentioned]).

Environment

Client: Web chat interface

Model: GPT‑5.5/5.4 (with “extra high” reasoning)

Account type: Free (also replicated on Plus/Pro)

Plugins/sub‑agents: None

Number of threads: 5 (new account)

Additional information

The bug appears to be a recursive addition loop in the token metering service—possibly triggered by an auto‑continuation or internal reasoning step. The same impossible numbers (7.2B, 3M) repeat identically across sessions and even after account resets, which strongly suggests corrupted server‑side state rather than a genuine usage calculation. We request an immediate audit of the raw billing logs, not just the front‑end counter.