To everyone complaining about usage...
Heat trend
Collecting trend data
The percentage is based on available heat signal, not comment count or independent people.
OpenAI model activity is surfacing — worth tracking for capability changes, ecosystem impact, and availability.
Large Language Models (LLMs) consume more tokens the longer a session runs, as tokens are used for inputs, outputs, and reviewing the context window for each new output. To manage token usage, users should compact sessions when they reach 40%-50% of their limit, compiling them to a new session to free up context window space. Models like GPT-5.6 Sol and GPT-5.6 Terra have a context window of 1,050,000 and a max output of 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