← sources

The models are approaching good enough.....price, for a bit will matter. But speed is going to take the win

source: reddit (reddit:ClaudeCode)published 2026-09-30 10:46ingested 2026-10-01 02:44

retained text

I run three 20x accounts. 2 Claude, and 1 Codex (that i fired up to see what all the fuss is about when Fable limits made it pretty useless and opus 5.0 was all we had). \[Codex account won't be being renewed for now\].....

I have big workloads. I fire up orchestration agents and then parallel work is done as soon as my tokens come back online for the week. I tend to blow my accounts for the week within 1 to 2 days of usage - hold on to the odd 3-5% per account to get final things done mid week as i review features, content that has been built. Opus 5.5 is stretching out my usage to about 2 days now out of the 5. Definitely a huge improvement. I have all of them working on the same code bases, so their performance is comparable.

What i have noticed is that

\- all of them are getting the jobs done. I no longer care at all which "system" i am throwing at the job - either Astra orchestrating or Fable, now Opus 5.5 orchestrating. I don't bother with the other details myself anymore. No point - time better spent reviewing the output and crafting the next steps.

\- Opus 5.5 is definitely quicker and a little less re-work is needed (done by itself). Astra - a little more re-work, but about 15%, and it identifies the need and takes it on itself, but less efficient because of it.

\- The ChatGPT models are still so much nicer to "talk with", however good Opus 5.5 is - still about 5x the bla bla that Chat's models produce.

\- Speed and efficiency are now the major differentiators. Astra is slower and dramatically less efficient from a token useage point of view, despite having no 5 hour limit. Maybe Sol 6.1 will even that score.

\- I have noticed that I completely no longer care at all which models are working on the code or the content. They are both going to make dramatic progress. Key is they have crossed my trust threshold.

\- Claude is doing the equivalent of about 2000-2500 hrs of a humans workload per weekly token budget. These are the figures that the model is giving back to me when i am asking it to estimate and it feels about right from looking at how it is considering it. That is about the equivalent of 1 year's worth of work with no holidays (5 days a week 8+ hours a day). Haven't asked Astra - going to be a bit less....but still, stepping back ----wow.

The old AI adage - *this* is the WORST it is ever going to be - rings loudd (I add the dobule dd to show i have hand crafted this message,... though it might be a double bluff).

These two major model-groups for me are essentially already doing everything i want. I could give them 10x the workload if i had the tokens and that is going to be the only thing that is going to decide who wins going forward. Any issues around long running task management, per token intelligence etc., are rapidly moving towards those becoming non-questions, as the good-enough  is starting to be reached.

What is left? - cost, and speed. Give me a model-group that is twice as good as today (what ever that means), and another that is at today's capabilities but cheaper and faster and I now would take that every time. That was why i tried Codex. As Astra > Opus 5.0....just turned out it was terrible at token use.

It feels to me that the good enough is starting to converge and compute access is going to rule. Who has access to the most compute long term? Elon probably, and the Grok models are accelerating. Where do the Chinese models play in all this? - not sure as they are, for quite a while, going to be compute constrained, but reaching that good-enough threshold is imminent.

The main hole in my line of thought that i can see is that 'twice (or x) as good' may mean something that i can't even fathom at the moment that changes the game. I am going to guess that the compute thing is the thing we see when the models start to nerf, as the labs switch back to using their compute to train the next...release a new model, keep the compute available all good. Whoever can keep consistent compute at the good enough level to the customer wins - e.g. at the Opus 5.5 level as of say yesterday when it was running at Peak-Opus 5.5.

document fetched 2026-10-01 00:44 · canonical https://reddit.com/r/ClaudeCode/comments/1wu1ov2/the_models_are_approaching_good_enoughprice_for_a