Why OpenAI Reset Codex and ChatGPT Work Quotas on August 29

OpenAI’s August 29, 2026 reset was primarily a remediation measure for paid Codex and ChatGPT Work users whose allowances could be depleted by inefficient product behavior or background processing. It restored applicable subscription usage automatically rather than requiring users to file individual claims. This was separate from API billing, where usage is metered by tokens and priced independently. 4 17

The reset did not mean that every subscriber received a permanent 10%–50% larger quota. OpenAI-related reporting said the fixes could make an unchanged allowance last 10%–50% longer, depending on how each person used Codex or ChatGPT Work. 8 4

What was consuming quota unexpectedly?

OpenAI’s engineering team initially identified three usage drains:

• Image-heavy context compaction: Long sessions containing images could become less efficient after repeated context compression. A related changelog entry says retained images are now counted against the context budget and older images are trimmed when necessary. 15 30
• Computer History overhead: Usage at the high end of Computer History activity was reported as unusually expensive. 15
• Automatic conversation titles: A title-generation feature was using more allowance than intended, creating model activity unrelated to the user’s primary task. 15

Later reporting described a broader set of eight fixes. The additional issues included agents continuing after they had effectively completed their goals, subagents being routed to more expensive models, and scheduled automations running more often or under conditions users had not intended. 4 8

The practical problem was invisible consumption. Reprocessing images during compaction could add repeated vision-related work; an agent that failed to stop could continue making model and tool calls; and background features could spend part of a user’s allowance without an obvious new request. These mechanisms explain why some users saw their quota fall faster than the visible amount of coding they had asked for. However, the available public material does not provide an authoritative, item-by-item breakdown of all eight fixes or a precise savings percentage for each one.

How much longer could the allowance last?

The reported improvement was 10%–50% more useful work from the same allowance, not a matching increase in the numerical quota. The range reflects workload differences. Users running long sessions with many images, Computer History, multi-agent workflows, or scheduled tasks were more likely to benefit than people using short, text-only coding sessions. 8

The image-related fix is supported by the published changelog, but no reliable public source here establishes an exact reduction in image-processing usage across all accounts. Users who did not encounter the affected behavior may therefore see little practical change after the reset.

What did Thibault Sottiaux say before the reset?

On August 24, Codex and ChatGPT engineering lead Thibault Sottiaux said the team had found the three initial drains and was using a “tiger team” to investigate reports and ship fixes. 15 Subsequent reporting said the team had worked through thousands of user reports before the broader August 29 reset. 8

That sequence matters: the reset was presented as compensation while OpenAI corrected the underlying efficiency and accounting problems, rather than as evidence that subscription quotas had been permanently raised.

Why the timing mattered as Codex adoption surged

Codex and ChatGPT Work were expanding quickly during the same period. Secondary reporting put combined usage at approximately 8 million users in mid-July, 15 million in mid-August, and 20 million by August 21. 10 23

The products were also operating alongside the GPT-5.6 Sol rollout and lower Sol pricing. Reuters reported that Sol’s standard short-context API price had fallen to $4 per million input tokens and $20 per million output tokens after an August 21 reduction. 17 OpenAI’s own model page also described GPT-5.6 as available across ChatGPT, Codex, and the API, with usage and pricing varying by product. 28

Rapid growth amplifies small inefficiencies. A feature that wastes a modest amount of compute in one session can become a major capacity, billing, and customer-support problem when millions of people are using agentic coding tools.

Why users saw repeated resets

Codex and ChatGPT Work had already seen several resets and limit changes during June, July, and August, including a reported global reset on August 11. 18 Earlier resets were connected to milestone celebrations, temporary policy changes, or efforts to keep users working while OpenAI adjusted the system. 14 23

It is not sufficiently supported here to confirm the specific claim that there were “six resets in two weeks” in July, or to attribute every reset solely to infrastructure failures. The stronger conclusion is that OpenAI was repeatedly calibrating quota metering, model routing, background features, and capacity controls while demand was rising.

That history also explains user skepticism. A previous 20-million-user milestone credit was described as a banked reset that users could redeem later, rather than an immediate refill of the active meter. 13 10 A banked credit can look like a missing reset if users expect their visible allowance to jump immediately.

How the reset fits with Plus and Pro limits

The reset came alongside a broader effort to manage Codex and ChatGPT Work capacity. OpenAI restored a five-hour rolling limit for ChatGPT Plus users after a period in which weekly limits were the main constraint. 5 6 Reports at the time said higher-priced Pro plans remained exempt, although future limits were possible as usage increased. 12

That policy tension reflects the economics of agentic coding. A single task can involve long contexts, tool calls, image processing, background activity, and multiple subagents. Fixed-price subscriptions are easier for customers to understand, but they make transparent and accurate usage accounting especially important when workloads vary so widely.

What about a usage-spending display?

OpenAI has described tools for reviewing activity and credit usage in ChatGPT Work and Codex, including ways for administrators to see adoption, approach-to-limit signals, and spending requests. 49 The available evidence does not establish the exact design or launch status of a broader user-facing spending display.

A detailed meter showing usage by model, task, images, agents, and automations would make quota consumption easier to understand, but that interpretation should not be treated as a confirmed product specification. For now, the clearest verified takeaway is narrower: OpenAI reset paid subscription allowances after fixing several reported sources of unintended consumption, while the long-term question remains whether agentic coding can be made predictable enough for fixed-price plans.