Codex Resets
308 points
• 2 days ago
• Article
Link
对于 OpenAI 的 Codex 和 ChatGPT Work 用户来说，追踪使用限额已经变成一场非正式的等待游戏，大家都盯着社交账号 @thsottiaux 的更新。没有官方的清除时间表或公开的更新日志；社区完全依赖他零星发布的公告，这些公告被跟踪并存档，构成了对这些偶尔"赦免"时间点的记录。
这些重置为经常触及周或小时上限的高频用户带来一种反复出现但难以预测的缓解。截至目前共有 35 次重置，平均间隔约 8.9 天；这些公告通常由真实的服务问题、系统迁移或庆祝性里程碑触发。在许多情况下，重置被用来补偿技术故障、延迟或导致使用令牌消耗异常加快的意外错误。
除了用于故障补偿外，重置也常被当作社区参与和激励的手段：当开发者达到 700 万或 800 万活跃用户等里程碑，或为插件等新功能上线庆祝时，会发放额外的使用额度。有些重置甚至被包装成季节性礼物，例如节日限额扩展或为成功的模型研究举行的庆祝活动。如今用户还可以通过桌面应用或网页版自行申请已累积的重置额度，不再仅仅依赖系统统一自动清除。
总之，这些更新凸显了管理高流量 AI 基础设施的脆弱性。相关信息显示团队以高速迭代为重，常把系统推到极限以支撑数百万并发用户。尽管"重置"作为一种非官方且混乱的做法仍存在，但它已经成为 Codex 文化的一部分，将技术维护转化为开发者社区的周期性庆典。
For users of OpenAI's Codex and ChatGPT Work, keeping track of usage limits has turned into an informal waiting game centered around the social media updates of @thsottiaux. There is no official schedule or published changelog for when these rate limits are cleared. Instead, the community relies entirely on his sporadic announcements, which are tracked and archived to provide a timeline of these occasional "blessings."
The resets serve as a recurring, albeit unpredictable, relief for power users who frequently hit their weekly or hourly caps. With a record of 35 resets and an average interval of roughly 8.9 days between them, these announcements are often triggered by a mix of genuine service issues, system migrations, or celebratory milestones. In many cases, a reset is used as a tool to compensate for technical outages, latency problems, or unexpected errors that cause usage tokens to drain faster than anticipated.
Beyond mere troubleshooting, these resets often function as community engagement tactics. Developers have been granted additional usage as rewards for reaching milestones like hitting 7 or 8 million active users, or to celebrate the launch of new features like plugins. Some resets are even framed as seasonal gifts, such as holiday-themed limit expansions or celebratory gestures for successful model research. The process has evolved to the point where users can now apply banked resets themselves through the desktop app or web interface, moving beyond simple automatic system-wide clears.
Ultimately, these updates highlight the volatile nature of managing high-traffic AI infrastructure. The messaging suggests a team focused on iterating at high speed, often pushing systems to their breaking point to support millions of concurrent users. While the practice of "resetting" remains an unofficial and chaotic part of the user experience, it has become a staple of the Codex culture, effectively transforming technical maintenance into a recurring celebration for the developer community.
200 comments • Comments Link
• 频繁重置使用额度并取消严格限额，使用户养成了更高频、更依赖 agent 的工作流。人们越来越担心这种"新常态"不可持续，一旦恢复到原先的限制，用户会明显感到降级，从而陷入潜在陷阱。
• 这些慷慨的重置普遍被解读为对新兴竞争对手的回应。尽管当前对用户有利，但普遍预期一旦确立市场主导地位或公司走向 IPO，价格将大幅上涨或访问会被严格限制。
• AI 定价中的"逐底竞争"因能力并非静止而变得复杂。尽管 token 作为大宗商品的价格可能下降，服务提供商却通过引入更新、更高效或更大规模的模型，提高成本与资源需求，从而"抬高底线"以确保盈利。
• 这些工具的企业级使用量激增，部分用户消耗了海量算力。公司通常把这笔开销与雇佣额外工程师的成本进行比较，将 AI 视为一种可以按需缩减的灵活可变费用，从而证明支出的合理性。
• 用于推广这些平台的增长黑客策略（如合并应用和频繁重置）旨在在 IPO 等重大财务事件前最大化用户获取与注意力。用户越来越意识到自己正在被投资者补贴，并将当前环境视为暂时的"赌场"时代。
• Anthropic 和 OpenAI 在策略上存在差异：OpenAI 更积极利用重置来建立好感并让开发者习惯其生态，而 Anthropic 更为限制，且限额政策有时模糊不一致，偶尔会疏远高级用户。
• 用户越来越多地自建遥测与监控工具来追踪使用情况，因为官方平台的报告常常含糊或具有误导性。这种技术性的"耍花招"文化凸显了人们对当前限额管理的挫败感，以及对最大化受补贴算力的迫切需求。
• 除了纯粹的 token 成本外，缓存效率、模型参数规模和 tokenizer 的"密度"等因素也会导致实际经济价值的巨大差异。随着平台引入复杂的门控和优化手段，用"每项任务成本"而非简单的单位定价来评估服务变得至关重要。
• 美国前沿实验室与不断崛起的国际替代者（如 Kimi 或 DeepSeek）之间的竞争正在催生一场"福利争夺战"。这种竞争短期内确实给用户带来了实惠，但许多人怀疑这是一场暂时的泡沫，一旦立法或市场整合开始，泡沫就会破裂。
• 对这些工具的依赖正成为一种重大的职业风险。随着用户将 AI 融入关键的编码工作流，潜在的"rug pulls"（即供应商突然提高价格或限制用户当前所依赖模型的访问）已成为真正的职业焦虑来源。
当前 AI 开发者工具格局的特征是一场在重大公司里程碑到来前抢占市场份额并确保高估值所驱动的激烈且不可持续的补贴战。虽然这为开发者创造了廉价访问和快速试验的"黄金时代"，但它也人为推高了使用模式，掩盖了推理的真实成本。在这种过剩带来的直接好处与开发者体验可能只是暂时便利、并会在市场整合或资金枯竭时消失之间，存在着明显的张力。 • Frequent usage limit resets and the removal of strict caps are conditioning users to adopt higher, more agent-heavy workflows. There is growing concern that this "new normal" is unsustainable and that reverting to original limits would feel like a significant downgrade, creating a potential trap for users.
• These generous resets are widely interpreted as a competitive response to emerging rivals. While currently beneficial for users, the expectation is that once market dominance is established or companies move toward an IPO, prices will rise sharply or access will be severely restricted.
• The "race to the bottom" in AI pricing is complicated by the fact that capabilities are not static. While commodity prices for tokens may drop, providers are effectively "raising the bottom" by increasing the cost and resource demands of newer, more efficient, or larger models, ensuring they remain profitable.
• Enterprise-level usage of these tools is skyrocketing, with some users burning through massive amounts of compute. Companies are often justifying these costs by comparing them to the price of hiring additional engineering staff, viewing AI as a flexible, variable expense that can be scaled down if necessary.
• The growth hacking strategies used to promote these platforms—such as merging apps and offering frequent resets—are designed to maximize user acquisition and "mindshare" before major financial events like IPOs. Users are increasingly aware that they are being subsidized by investors, treating the current environment like a temporary "casino" era.
• Anthropic and OpenAI differ in their strategies, with OpenAI appearing more aggressive in using resets to build goodwill and habituate developers to their ecosystem. Anthropic's more restrictive approach and opaque, inconsistent limit policies have occasionally alienated power users.
• Users are increasingly building custom telemetry and monitoring tools to track their usage, as official platform reporting is often vague or misleading. This technical "shenanigan" culture highlights the frustration with current limit management and the desperation to maximize access to subsidized compute.
• Beyond pure token costs, factors like caching efficiency, model parameter counts, and the "density" of tokenizers create significant variations in actual economic value. Evaluating services by "cost per task" rather than simple unit pricing is becoming essential as platforms introduce complex gating and optimization tricks.
• The competitive tension between American frontier labs and rising international alternatives (like Kimi or DeepSeek) is driving a "perk war." This competition provides tangible benefits to users, though many suspect it is a temporary bubble that will burst once legislative or market consolidation takes hold.
• Dependency on these tools is becoming a significant professional risk. As users integrate AI into critical coding workflows, the potential for "rug pulls"—where providers spike prices or limit access to the models users now rely on—is a source of genuine professional anxiety.
The current landscape of AI developer tools is characterized by an intense, unsustainable subsidy war driven by the need to capture market share and secure high valuations ahead of major corporate milestones. While this creates a "golden age" of cheap access and rapid experimentation for developers, it is also artificially inflating usage patterns and masking the true costs of inference. There is a palpable tension between the immediate, tangible benefits of this excess and the underlying fear that the current developer experience is a temporary convenience that will evaporate once the market consolidates or the funding dries up.