https://help.openai.com/en/articles/20001625-understanding-chatgpt-and-codex-usage-limits
Starting in December, Plus, Pro, and Business plans’ usage allowance will be shared across ChatGPT and Codex.
OpenAI confirms that:
- ChatGPT and Codex will share a token-based allowance.
- Instant chat will not consume this allowance.
- More complex tasks and advanced models will consume more tokens.
- Users can check their remaining allowance under Settings → Usage.
- Users can purchase extra credits if they exhaust their allowance.
TL;DR - ChatGPT conversations that use advanced reasoning will compete with Codex for the same allowance. And it will be token based, rather then message / turn based. How many tokens? Fuck you, that’s how many. Want a coder and chat with a non brain dead LLM? Pick one.
So, if you’re thinking of re-upping your OpenCode Go or OpenRouter access (or you’re fortunate enough to have a rig that can run Qwen3.8 27B, Flash or the soon to be arriving Qwen 4’s), now’s the time to fuck off jippity.
Personally, I’m going to ditch it for OR credits and selective use of GLM, Kimi etc for code review. I don’t have the rig for a local GLM :(
Ride was always going to end but fortunately, open weight models are a thing.
Now if I could only run Qwen3.8-27b at not minutes-per-token…
I’d guess that chat sessions use a trivial amount of tokens in relative terms.
I’m finding the crack down on security stuff is making me go more local.
I’m also finding the newest models regressing a lot, they burn a lot of tokens for pretty shitty results.
Excited for qwen 4 though.
It’s a two pronged attack for sure. Make local expensive, make cloud worse.
I think there’s still some wiggle room to be had in the 8 to 16GB VRAM range…so I’m hoping Qwen 4 series brings its A game.
Qwen 125bb flash next is probably one of the best models I can run (strix halo), it’s fast and smart. But it burns too many tokens on reasoning even on low effort.
If they do a qwen4 125b a6b or whatever I will be happy. No idea if that’s what their qwen4 flash will be or not.
A Strix. In 2026. Must be nice :)
It is, however it’s still pretty slow and not many models find the sweet spot that Qwen did.
I feel most models would be better on a faster GPU as MoE sizes are too big right now. I can get comparable tok/s on large models compared to discrete GPUs but much slower on small sizes.
The curse and blessing of unified memory!
I was considering getting a BC-250 (a cut down PS5) with 16GB unified memory, but at 400GB/s bandwidth.
When they were cheap, it almost made sense. But realistically, I can get similar speeds with a MoE thanks to cpu and system ram, and 16GB is just not enough for anything top shelf (no, Prism MLs Bonsai models don’t count. I don’t care what the benchmarks say - they’re lobotomised).
Still, given the chance to get a Strix on the cheap, I’d go for it. There’s something very nice about a low power box that can run a 120B model at speed.
I am definitely happy with it.
One of the best things is I run some local services and I can have embedding models, decision models, and small LLMs always in memory for my hosted apps. Then it’s only a few seconds to load up a large LLM like qwen 125b or similar.
Kinda wish I’d gone in one those BC250s as well, there are good guides on them now. 400gb/s is pretty good.
Oh, your running some JEVs? How are they? I just can’t figure out a use for em.
Yes, concurrency is amazing. I can pull some tricks with CPU (OCR, TinyBERT, Klein etc) while the GPU runs inference…but it’s not a Strix.
Yeah, the BC-250 had / has real promise, esp with the core unlocks…but like P40s and P100s (server cards) people twigged on “this is cheap and fast”. I can’t get one locally for under $600 now. They use to be $150. By the time you fit it out properly, it’s nearly $1K. Not worth it.