OpenAI Doubles Pro Plan Quotas; Article Reviews Backend Token Consumption of AI Coding Tools

On October 29, OpenAI product lead Tibo announced that effective immediately, the usage limits for the Pro 100, Pro 200, Pro 500, and Business Premium plans would all be doubled.

There was also an important note: after listening to user feedback, OpenAI decided not to reduce the usage quota for Pro 200, so that plan's ChatGPT Work and Codex usage will remain at their original levels. At the same time, Dots will continue to enjoy expanded quotas for now, and its usage is independent of the account's default quota and is not counted against the default quota.

This adjustment effectively responds to the controversy sparked by the earlier restructuring of the Pro subscription system. OpenAI had previously launched the $500 Pro 500 and adjusted the benefits of the $200 Pro plan, and existing Pro 200 users had faced the prospect of some usage quotas being lowered. This announcement withdraws the related reduction arrangements, meaning that user feedback did indeed influence the final policy decision.

Subsequently, a user further asked whether Pro 500 means getting 50 times the usage of the Plus plan, and Tibo answered directly, "Yes." This also reflects that OpenAI is further widening the usage quota gap between different paid tiers, giving high-intensity AI coding and workflow users more room to work.

The community response was not uniformly positive. Some Pro users welcomed the quota expansion, but some $200 Pro users pointed out that they had already enjoyed relatively high quotas, and this time only the previously planned reduction was withdrawn, without any genuinely new benefits being added.

"I'm glad that users on other Pro tiers and lower-quota plans got real quota increases—those are the real optimizations. But packaging the cancellation of the originally planned quota reduction into the 28-day improvement plan and promoting it externally as an across-the-board doubling upgrade is misleading. Frankly, for users who have been following this matter and relying on official announcements for planning, this framing is really hard to accept," one user said.

At the same time, some Plus users expressed dissatisfaction that this adjustment did not cover lower-priced plans, believing that OpenAI is further widening the usage experience gap between different paid tiers.

In addition to raising quotas, Tibo also emphasized that OpenAI will continue to launch better models and efficiency optimizations so that users get more actual usage over time. Notably, this adjustment is Day 25 of OpenAI's 28-day product improvement plan. Tibo then joked that OpenAI is graduating from being a "quota reset company" to being a "token company."

Andrew Ambrosino, who is responsible for Codex and ChatGPT products at OpenAI, also said that the team has launched an intensive round of improvements focused on addressing app simplicity and performance issues, which he is personally overseeing.

This means OpenAI does not intend to rely solely on raising plan caps to meet user needs, but hopes to simultaneously improve model capabilities, inference efficiency, and output per unit of compute. For developers who increasingly rely on Codex, whether they can obtain stable and sufficient usage quotas has become an important factor in subscription decisions.

In addition to OpenAI, over the past year Anthropic has also frequently adjusted Claude Code's usage limits, even going through a process from tightening to expansion and then partial clawback.

In March 2026, Anthropic tightened the 5-hour usage quotas for free, Pro, and Max users during peak hours; after reaching a compute partnership with SpaceX in May, it doubled Claude Code's 5-hour quota and canceled peak-hour quota reductions for the Pro and Max plans. At the same time, the company also temporarily raised Claude Code's weekly quota by 50%, but after ending the promotion in September, it retained only a permanent quota 25% higher than the original level.

Anthropic's Max 5x and Max 20x plans are advertised as offering 5 times and 20 times the usage of the Pro plan, respectively, but this multiplier actually refers to usage within each 5-hour session window rather than cumulative weekly usage, and users are also subject to an overall weekly limit. In September of this year, a group of users filed a class-action lawsuit against Anthropic on this basis, arguing that the relevant limits were not sufficiently and prominently explained during the purchase process, resulting in a gap between the actual usage experience and purchase expectations.

1 Background agents consume a lot of tokens in inconspicuous places

The quota issue with AI coding tools is not simply a matter of plan pricing. As agents gradually take over more development tasks, background automatic execution, task scheduling, and multi-agent collaboration have also begun to consume large amounts of compute, and there are even cases where users have consumed a considerable amount of usage quota without directly perceiving the tasks.

In June of this year, OpenAI's Codex experienced a similar problem. At the time, many users found that usage quotas that originally took a week to consume were suddenly used up within one or two days. Some $200 Pro users said that quotas that previously required a full week of continuous work to use up could later be exhausted in a single day.

Tibo later explained that the problem was related to abnormal background workloads: automatic code review and auxiliary sub-agents sometimes ran more frequently than expected, even repeatedly executing tasks, or retrying excessively after errors occurred, causing unnecessary compute consumption. In addition, the Codex dashboard also displayed some activity that was not actually billed to users, further exacerbating users' confusion about abnormal quota consumption.

To address this, OpenAI organized an emergency investigation, set up a dedicated team to investigate the issue, deployed fixes and reset all users' usage quotas, while also increasing background usage monitoring so that similar problems could be detected earlier. Tibo also said that the team would provide users with an additional quota reset to make up for the impact of the previous abnormal consumption.

Anthropic's Claude Code was also found to have a similar problem, except that the consumption occurred in a more inconspicuous part of the process.

On August 15, a developer submitted an Issue in Anthropic's Claude Code GitHub repository, requesting that the official team provide a formal switch allowing users to turn off sub-agent status summaries without sacrificing the entire sub-agent feature.

The problem lies in Claude Code's background status refresh mechanism. When users start multiple sub-agents, the system periodically calls the model to generate a short status prompt for each active agent, such as what task is currently being executed. According to the developer's analysis of Claude Code's internal requests, this mechanism is triggered by default every 30 seconds, and each time it resubmits the sub-agent's complete conversation history and tool definitions to the model to generate a summary. The final output is usually only a few dozen tokens, mainly used to update the status text in the interface, while also helping users and Claude Code keep track of the progress of each background agent.

The developer performed packet capture analysis on a session lasting 4 hours with 27 sub-agents started. Of the 1,496 API requests, 339 were agent_summary, accounting for 22.7% of all requests; these requests generated about 62.76 million raw tokens, accounting for 22.2% of total raw tokens. After weighting according to Opus 5's cache read, cache write, and output prices, their cost still accounted for about 18% of the entire session.

Another developer also found a similar problem when testing with 4 sub-agents for 1 hour and 45 minutes. Among them, a long-running agent's context size reached about 380,000 tokens, and 103 status summary requests generated 21.8 million cache read tokens, accounting for 26% of the entire session's cache reads, while the final output was only about 2,100 tokens. According to the developer's estimate, this agent alone could generate about 45 million cache read tokens per hour just for generating status prompts.

What is more troublesome is that these background requests are set with skipTranscript: true and usually do not enter the regular session transcript file. Therefore, third-party tools that rely on session records to count consumption may underestimate actual token usage. What users see is a few constantly changing words in the agent status bar, but behind the scenes a large amount of context reading may have occurred.

On October 28, the developer "Flamerix AI" also publicly released its monitoring results. By analyzing Claude Code's OpenTelemetry telemetry data, it found that within a one-week statistical period, agent_summary accounted for 16.4% of all estimated costs, 19.7% of total API requests, and the ratio of input to output tokens was about 9000:1. In other words, to generate a small amount of status text, the system needs to repeatedly read a large amount of context.

Flamerix AI subsequently released an open-source script under the MIT License, attempting to intercept the relevant timers through Bun's preload mechanism and turn off status summaries without modifying the Claude Code binary. The script was tested on multiple versions in a Linux environment. After disabling summaries, background sub-agents, Forked Skills, and background sessions were still able to run normally; the only difference was that the Agents panel no longer continuously updated the status text. However, this approach depends on Claude Code's internal implementation and is not yet an officially supported solution from Anthropic, and future version updates may also render it ineffective.

As the matter continued to spark discussion in the community, an Anthropic engineer named Edwin Arbus, who claimed to be a member of the Claude Code team, also came forward to respond.

He explained that this mechanism only targets background agents that are actively working; idle agents will not continuously trigger summary requests. The 16.4% measured by Flamerix AI may come from teams that make heavy use of background agents and is not broadly representative. According to Anthropic's telemetry data from last week, among Claude Code users who submitted telemetry information, agent_summary accounted for about 2% of total cost at API list prices and slightly less than 4% of total requests; if only the terminal environment is considered, this proportion is about 6%.

However, Arbus also admitted that even 2% is still too high for a feature mainly used to generate status prompts. He revealed that Anthropic has already fixed the most expensive category of cases in Claude Code 2.1.311 and later versions, namely the problem of summary requests failing to hit the prompt cache. The team is currently developing a cheaper implementation that no longer submits the full conversation history each time, but instead sends brief notes and lets a smaller model generate the summary.

However, this does not mean that background summary overhead has been completely eliminated. On October 29, developers were still reproducing the relevant requests on Claude Code version 2.1.312. Another developer analyzed 129 background agent sessions within 30 days and found that agent_summary accounted for about 14% of requests and 10% of estimated costs, and in one session containing 40 agents it even reached 24%.

This also shows that the problem is not simply an excessive number of background requests, but that context size, cache hit rate, model selection, and refresh frequency jointly determine the final cost. For long-running agents, even if every request hits the cache, repeatedly reading hundreds of thousands of tokens of context still adds up to a considerable expense.

2 How can AI programming tools provide stable service?

Developers are willing to invest large budgets in AI programming, and the fundamental reason is still that it may bring real productivity gains. However, increased tool usage does not equal a proportional increase in benefits, and how to measure return on investment (ROI) is becoming a new problem.

Fixed monthly subscriptions and pay-as-you-go billing have different economic logic. For API businesses charged by token, the more users consume, the more revenue vendors usually receive; but for fixed monthly subscriptions such as Pro and Max, the price users pay is relatively fixed, while inference costs may continue to increase with usage. Especially when agents can run autonomously for long periods and call multiple sub-agents in parallel, a small number of heavy users may generate computing costs far higher than those of ordinary users.

Anthropic's infrastructure investment also illustrates this pressure. Its IPO filing also points out that computing power is Anthropic's largest cost item, and the company's computing power and infrastructure spending commitments to partners over the next 10 years exceed $600 billion, a considerable portion of which are contracts that cannot be canceled or do not depend on actual usage. For a vendor like this, computing power is both the foundation of product supply and a major cost that must be managed over the long term.

Therefore, tightening quotas can control per-user costs in the short term and protect the gross margin of the subscription business; increasing quotas helps attract heavy developers, enhance product stickiness, and expand usage by enterprise customers. But if vendors frequently change usage rights that users have already formed expectations around, it may damage brand trust and even prompt enterprises to establish multi-vendor strategies, reducing dependence on a single product.

What AI programming tools ultimately sell to developers is not just model capability, but a kind of R&D productivity that can be continuously invoked. If vendors cannot let users accurately predict costs, nor guarantee the continuous operation of critical workflows, then even if model capabilities are stronger, developers may choose to limit usage, diversify vendors, or even reassess procurement plans.

Reference links:

This article comes from the WeChat public account "InfoQ" (ID: infoqchina), author: Chu Xingjuan, published by 36Kr with authorization.