To everyone complaining about usage...

查看原文

OpenAI模型发布订阅权益

热度趋势

趋势数据积累中

百分比基于当前可用热度信号，而非评论数或独立用户人数。

推荐理由

OpenAI 相关模型动态已经出现，适合跟踪能力变化、生态影响和后续可用性。

AI 摘要

大型语言模型（LLMs）的会话运行时间越长，消耗的tokens就越多，因为tokens用于输入、输出以及审查每个新输出的上下文窗口。为了管理token使用量，用户应在会话达到其限制的40%-50%时进行压缩，将其编译到新会话中以释放上下文窗口空间。例如，GPT-5.6 Sol和GPT-5.6 Terra等模型的上下文窗口为1,050,000，最大输出为128,000。

This may be obvious, but for those who don't know... the longer you run a session, the more tokens you will use. LLMs use tokens for inputs, outputs and review the context window for every new output. The more session text it processes, the more tokens burn, the faster usage gets gobbled up.

Additionally LLMs get dumber the long you run a session. Every model has capacity constraints built in, and once you cross 40% of that limit, there is too much information the model has to process to maintain quality output.

Matt Pocock explains these limits really well here:

https://youtu.be/nKSk_TiR8YA

https://youtu.be/-uW5-TaVXu4

Here is a breakdown of the context window capacity and max output for each of the models available in Codex:

Codex model Context window Max output GPT-5.6 Sol 1,050,000 128,000 GPT-5.6 Terra 1,050,000 128,000 GPT-5.6 Luna 1,050,000 128,000 GPT-5.5 1,050,000 128,000 GPT-5.4 1,050,000 128,000 GPT-5.4 Mini 400,000 128,000 GPT-5.3-Codex-Spark Not publicly documented separately Not publicly documented separately If you are running into limits then you need to compact your sessions when you can. Once you reach 40% - 50% you should compile the session to hand it off to a new one to free up context window space.

Also note that for those of you who use the voice feature, you are likely speaking WAY more words than you would type, which means more words = more token usage = faster drops in capacity.

To solve for this I created a skill called $context-capacity that, when run, tells you how much context capacity you've used, how much you have left, and the cumulative session usage with a recommendation. Here is what that output looks like for one of my sessions:

Recommendation: Handoff

Current context load: 144,827 / 258,400 tokens (56.0%)

Estimated remaining capacity: 113,573 tokens (44.0%)

Cumulative session usage: 289,355 tokens — cumulative, not current occupancy

Confidence: Exact recorded metrics with derived capacity. The current load exceeds the skill’s 40% handoff threshold.

The website and promo-video handoffs already created are ready for separate sessions.

Here's a link to the skills for $context-capacity and $handoff for anyone who wants to use it:

https://github.com/marcushackler/codex-skills