[Bug] 3 ChatGPT Work travel-search prompts consumed ~73% of the 5-hour quota on GPT-5.6 Sol Medium

What issue are you seeing?

A single short conversational travel-search session in ChatGPT on the web using Work mode consumed approximately 73% of my 5-hour usage window after only three user messages.

This was not Codex Desktop or Codex CLI. The session was ChatGPT Web + Work.

This was not a software-development workload, repository audit, large code generation task, or intentionally long autonomous run. It was ordinary web/travel research: searching for inexpensive flights, adjusting a connection preference, and then asking for booking links for one date range.

The usage is dramatically higher than I would expect for this workload and looks similar to the quota-consumption/accounting regressions reported elsewhere around Codex/Work usage.

Environment

• Product/surface: ChatGPT Web + Work mode
• Not reproduced through: Codex Desktop / Codex CLI
• Subscription: ChatGPT Plus
• Model: GPT-5.6 Sol
• Reasoning: Medium
• 5-hour quota consumed by this one chat: ~73 percentage points
• User messages in the chat: 3
• Date observed: 2026-09-14

Reproduction / exact prompts

Geographic destinations and connection locations are intentionally redacted. The wording is otherwise preserved as closely as possible.

Prompt 1

Мне нужно в [REDACTED DESTINATION]. В промежутке с 5 октября до декабря. Нас двое. Мы в [REDACTED ORIGIN]. На весь отпуск мы выделяем 5-7 дней, не важно каких. В идеале так, чтобы обратно полететь через другую страну ([REDACTED COUNTRY]?). Ищи билеты. Приоритет - цена

English summary: Find cheap flights for two people sometime between October 5 and December, for a 5–7 day trip, ideally returning via another country. Price is the priority.

Prompt 2

а если хочется пересадка в [REDACTED CONNECTION CITY] без ночёвки? 2. а если октябрь?

English summary: What if the connection is through a specific city without an overnight stay? And what about October?

Prompt 3

дай ссылки и всю информацию о бронировании на 02.11 - 08.11. передаб более умному

English summary: Give me the booking links and all booking information for Nov 2–8; hand it off to a smarter model if needed.

Actual behavior

Across these three messages in one ChatGPT Web + Work chat, the 5-hour usage meter dropped by approximately 73%.

The workload was normal consumer travel research. Even allowing for multiple web searches and itinerary comparisons, consuming nearly three quarters of the entire 5-hour allowance for three messages appears disproportionate.

Expected behavior

Quota consumption should be reasonably proportional and predictable for ordinary research workloads. Three travel-search prompts should not consume ~73% of a 5-hour Plus allowance unless there is an unusually large amount of model work occurring, and if that is expected behavior it should be observable/explainable to the user.

Related reports

The observed quota behavior appears similar to reports such as:

• #41957 — comparable tasks showing ~9x quota-consumption differences
• #45073 — ~86% of the 5-hour quota consumed in ~26 minutes with only two prompts on GPT-5.6 Sol Medium
• #41220 — tracker for abnormal quota depletion/accounting inconsistencies
• #44213 / #44233 — server-reported quota/accounting discontinuities

Those reports may use different Codex surfaces; this report is specifically about ChatGPT Web + Work mode.

I can provide additional session telemetry if there is a recommended way to export a sanitized usage trace for this specific Work chat.

Start with related reports #41957, #45073, #41220, #44213, and #44233, then determine whether ChatGPT Web + Work telemetry is in scope for this repository. Use the three reported prompts and the approximately 73% quota drop as the reproduction; done means explaining or correcting the quota accounting and verifying the usage meter behaves proportionally.