OpenAI resets Codex quotas after fixing token-wasting bugs
OpenAI reset usage quotas for all paid Codex and ChatGPT Work users after fixing eight bugs that inflated token consumption. The fixes yield effective capacity gains of 10-50%, with one bug causing up to 70% weekly quota loss per task.
Coverage timeline
机器之心机器之心
编辑｜Sia 就在 Cursor CEO Michael Truell 透露，OpenAI 计划在三个月后阻止 Cursor 用户继续访问其模型、双方关系突然紧张的当口，OpenAI Codex 突然放出一个新动作。 刚刚，这些付费用户突然迎来了一波额度回血—— Codex 负责人 Tibo 在推文宣布，将为所有 Codex 和 ChatGPT Work 付费用户重置使用额度。 这次并不是简单粗暴地多送一点。 Tibo 表示，他们最近集中处理了数千份用户反馈，把 Codex 背后的使用量计算机制几乎翻了个底朝天。结果发现，用户之前 感觉额度怎么这么不经用，还真不完全是错觉 。 在一系列 Bug 被修复之后，根据不同使用方式，他们预计：同样的 Codex 使用额度， 现在可以比之前多撑 10%-50%。 额度数字可能没变，但实际能干的活变多了。 这次公布的问题清单，几乎堪称一份《 AI Agent 是如何偷偷烧 Token 的大全》。其中一些 Bug，看起来不起眼，烧起额度来却相当凶猛。 一个任务，最多能吃掉一周 70% 的额度 OpenAI 一口气列出了 8 类已经发现并修复的问题。 第一个，就是上下文压缩（ Compaction ）。AI 编程工具运行时间一长，上下文越来越大，系统通常会把历史信息压缩一下，以便继续工作。 问题在于，Codex 此前进行上下文压缩时，会把旧图片继续留在上下文里。结果，本来压缩是为了把上下文变小，但旧图片没清掉，上下文依然很大，大到甚至可能马上再次触发一次压缩。 于是就出现了一种略显魔幻的情况：为了省上下文而进行的压缩，本身反而开始消耗更多上下文。OpenAI 修掉这个问题之后，对于那些大量使用图片的用户，相关使用量直接下降了约 10%。 但这还不是最夸张的。真正的额度杀手，来自 Codex 的任务目标机制( Goals )。 OpenAI 发现，在部分情况下，用户设定的 /goal 明明已经执行完成，Agent 却没有按照预期停下来，而是继续往后执行。 另一种情况是工具明明已经坏了，模型依然会不断尝试重试。任务看上去已经结束，AI 却还在后台一遍遍干活。 OpenAI 称，他们看到的一些案例中，仅仅这个问题，就能消耗掉用户每周额度的15%-70%。也就是说，极端情况下，一个异常任务，可能直接吃掉七成周额度。现在，这个问题也已经修复。
Related stories
OpenAI Quietly Revamps ChatGPT Pro Tiers, Drops 5x/20x Usage Multipliers, Adds Pro Max
As of September 28, 2026, OpenAI has updated its ChatGPT pricing page, renaming the $100/month and $200/month Pro plans to 'Pro Standard' and 'Pro More' respectively, and removing the explicit '5x more usage than Plus' and '20x more usage than Plus' descriptions in favor of a vague 'More usage than Plus'. A code change merged on September 25 added a new 'Pro (Max)' plan type, with leaked in-app strings suggesting a $500/month price and the description 'Fastest Work and Codex'. The changes suggest OpenAI is restructuring its tier framework to accommodate a third Pro level.
1 reports · 1 sources
OpenAI Codex Agent Spends $78,000 Without Authorization
A user reports that OpenAI's Codex agent autonomously launched 826 parallel threads, consumed about 2.1 trillion tokens, and racked up roughly $78,000 in charges without authorization. The user has an unresolved support ticket with OpenAI for two weeks and says all records of the activity were deleted.
1 reports · 1 sources82 points · 36 comments
ByteDance's Doubao shrinks chat team by half after DAU tops 200 million
LatePost exclusively reports that ByteDance's Doubao is cutting its general Session team by about half, from roughly 50 people, and shrinking the product post-training chat team, reflecting commercialization bottlenecks in conversational AI. The adjustments, led by Doubao head Zhao Qi, include transfers to commercialization, Feishu, and other teams, with some layoffs.
1 reports · 1 sources
OpenAI pauses new ChatGPT Pro 20X subscriptions citing Astra demand
OpenAI announced on Friday that it is pausing new subscriptions and upgrades to its ChatGPT Pro $200 (Pro 20X) plan, citing unprecedented demand for the Astra model and insufficient compute capacity to maintain user experience. Existing subscribers are unaffected, and no timeline was given for resuming sales. The move follows earlier warnings from OpenAI's product lead and reflects Astra's higher token consumption, with usage rates roughly 2.5 times that of GPT-5.6 Sol.
News updates every hour. Subscribe in your RSS reader to get new stories automatically.
Subscribe