OpenAI made a change to Codex that most people scrolled past. For anyone who works in long agent sessions, it's worth more than the last few model releases. For logged-in Codex and Work users, the prompt cache now never expires. That sounds like a small technical detail, but it decides how far your subscription goes. Here's why. What the cache does Each time you send a message to a coding agent, the model reads the whole conversation again, not just your new message. That includes the system prompt, tool definitions, repo context, and every earlier turn. In a real coding session, that earlier context is usually over 90% of the input tokens. Your new message is the small remainder. The cache lets the provider skip reprocessing context it has already seen. Cached tokens cost about 5-10% of the normal input price. What used to happen The cache expired after a while. If you stepped away for lunch, a meeting, or the night, your next message paid full price on everything that came before. Here's an example. Say your thread holds 150k tokens of context and you add 5k new ones. With a warm cache, the 150k cost about the same as 7.5k to 15k normal tokens. With the 5k new ones, the message cost the equivalent of 12.5k to 20k. With an expired cache, you paid for all 155k at full price. The same message cost 8 to 12 times more just because you'd taken a break. This happened on the first message after every pause, so a thread you came back to three times a day paid it three times a day. Nothing in the interface showed this. Your limits just ran out sooner than you expected. What changed For logged-in Codex and Work users, that cost is gone. Come back after an hour, a day, or a week, and everything before your new message is a guaranteed cache hit. You pay the normal rate only on the tokens you add. What didn't change The API still caps cache retention at 12 hours (new), and cache writes there cost more. The permanent, guaranteed hit is only for the subscription products. The model also isn't any smarter. It has no new capabilities. Long context still makes each new turn cost more than a short one would. It's just roughly 10 times cheaper than it was after a pause. People who keep one thread per project and come back to it for days. People who pause often to review diffs, test, or think, then pick up where they left off. People who used to start fresh threads to save on limits and lost their context in the process. That's a win. If that's how you work, your subscription now covers much more work at the same price. Cache hit rate was always the main thing deciding whether a subscription's limits felt generous or tight.

I guess I'm a night owl with @t3dotcodes Those 2–4 a.m. nights… this is not good. I should sleep

I'm really not liking using GPT 6.1 SOL. It gives me headaches. Something I just ran into: -"We have to do this, that, the other thing, and X in TLDRAW OFFLINE. BUT FIRST you need to update tldraw and its skills." -And GPT 6.1 SOL started working without updating it. -So I steered it and said "bro, I told you to update it first before you start working," -Then it updated it, said "Done, updated," and ended the turn right there. -Now I have to tell it "Bro, if it's already updated, then go do what you were supposed to do." This kind of behavior sometimes makes you want to break something in the setup. Always back to Opus 5.5 (Sometimes a little bit stupid too, but less than GPT 6.1 SOL or Astra. Here I really agree with how @theo frames his chart of intelligence "spikes" and the stability of answer quality. OpenAI's models aren't stable.)

I need a Claude Code reset

This is it

This will be the most beautiful livestream of all time. Maybe even better than the Artemis II lunar flyby videos.

the difference is brutal

Does it need a practical use? Just enjoy the engineering, man. It’s a luxury item, a supercar. You don’t buy a Bugatti (or a Roadster that can flip and hover) because you’re going to drive 500 km/h or fly every single day. You buy it because it’s beautiful, cool, and because you love the engineering behind it. No one said its a daily driver for everyone, literally a super-car. Enjoy life.

Whos gonna tell him

THATS ROCKET LEAGUE IRL WTF TESLA