Comment by madrox
13 hours ago
I am finding that I am now less interested in better models than I am in token budgets. My issue with Anthropic models now is that I don't feel like I can rely on them as a daily driver because they'll dry up before my quota resets.
I am becoming dependent on AI to make a living, and I need predictable spend on it. If I know I can't use a model regularly all month, my enthusiasm is limited.
I urge Anthropic to get better at this aspect of their business so I can come back to it.
> I am becoming dependent on AI to make a living
IMO, if you depend on AI to make a living, I'd invest in hardware for local inference, and learn on how to effectively make a living using AI inference you control, on hardware you control. Sure, economically speaking it's way cheaper to use one of these heavily subsidised services (for now), and their models are faster and more capable, but if your livelihood depends on AI inference, and you are renting AI inference, you are a being a serf of the tokenlord. And your livelihood depends on the whims of the tokenlord. They can increase rent prices, they can decide you can no longer do whatever you are doing, and you have no recourse, because you are dependant on them to make a living.
There are a lot of things in my toolchain pre-AI that I did not own and relied on to make a living. Mobile developers are in even worse shape, and iOS developers doubly so. The idea we were somehow less beholden before AI, I think, is silly.
None of us can wholly do our trades without support. Local inference is a fun idea, but you'll be out-competed by the serfs, as you call them.
Except there's a huge gulf of self-hosting and using API hosts - no way you can reach the economics of a shared host. Privacy is a problem but you can chose who you host with and where it's hosted (which jurisdiction).
When privacy/compliance really starts to matter it's up to the client/business to provide you with tooling - you're not running that on your own hardware anyway.
So the local AI for individuals is just a hobby/gimmick at this point not a rational decision. Self-hosting for business is a different story.
I'm not sure. The problem with the cloud llm's is they are complete black boxes that change frequently and randomly day by day.
If you run Qwen 3.8 on your own hardware, every single day, it's the exact same model running in the exact same way.
Yes, it's no where near as "smart" as the cloud based models. But it's consistent.
So the workflows/processes/ways of working you create will work mostly similar day to day.
With Claude/OpenAI you frequently find days where the models are useless, and days when they are out of this world.
So I guess the choice comes down to:
1. Randomly the smartest thing on the planet with unpredictable rate limits that is mostly amazing, but frequently messes with your workflows
2. A really good local coding model that is consistent every day with no rate limits
I'm not sure. My gut feeling is maybe the right answer is a mix of both.
Gambling on the biggest models, hoping they are working smart that day, when planning or doing very complex work. Then doing most of the tasks/daily work using local models??
I literally only make it halfway through the week until my weekly usage runs out. This is using only Opus, no fable, and I'm on the max x20 plan. It's become ridiculous.
These comparisons are meaningless
I use Opus every day and easily have most of my weekly limit left over at the end of the week
just buy two 20x, no?
or switch to codex
Invest a bit of your time into optimising usage cost. Anthropic has first class docs, actually read it or ask llm to read them all for you and summarise most important points / ask to to reflect it on your .md files. Maybe silly thing like dropping your default thinking effort by one level or adding (sub)agent pinned to other model is all there it to completely fix it or maybe you have instructions that encourage big dumps in CLAUDE.md/AGENTS.md that needs splitting so progressive disclosure works correctly? Naively sending everything to the most expensive model on high thinking effort is anti pattern and will drain quota quickly.
My personal guess is that it's one of those. With effective context engineering it's hard to use all 20x quota, the limit becomes your own attention and time really.
You may argue that you're doing multiple, parallel extreme effort tasks – which may be true but then again, there will be results to actually look at sooner or later and that takes time.
Agreed. The area I think will become more prevalent in the future for organizations are cost per intelligence -- effectively efficiency. An unoptimized model that costs 90x more than another that is only 10-15% less intelligent is something I would say is not a good deal.
I am on the Claude Max 20x plan, and this still happens when using Fable 5/Opus 5. I would run out of weekly quota in 2 days, whereas Opus 4.8 would last the entire week, and sit at about 80-90% at the end.
GLM 5.3-flash fits the bill
What is it equivalent to?
What kind of things are you using it for?
I haven't tested it yet but on all the benchmarks it looks like it's 5-7x slower for agentic tasks.
I made some webapps with it, and have it running my hermes agent (which also does a lot of coding, but not webapps).
Not sure what it's equivalent to, but it's super cheap and I am happy with the results
It's a mix of slightly worse kimi k3 for UI work and slightly smarter than luna for everything else.
But yeah, it's very slow. I've put it to work as an LLM-as-RAG agent.
I'm with you, for what I usually do most models are already more than enough.
What I'm really keen on is better auto-reasoning so I don't have to constantly have the constant inner debate on which reasoning effort to pick for each task.
I seriously hate the none-low-medium-high-xhigh-max-ultra etc that we have now, with companies frequently recommending different ones on each new model release, etc.
It's apparently called Adaptive Test-Time Compute or Dynamic Test-Time Compute and companies are apparently working on it (according to some LLM :shrug:)
Adaptive reasoning is known to be an extremely hard problem to solve, though. It requires you to predict whether a certain LLM, with a certain effort level, with a certain prompt, will give you the right answer.
This feels like exactly the kind of problem domain that belongs in (and can be solved by) RL?
Try gpt 5.6 Luna max
overthinks, been slow lately through the official api (slower than glm 5.3 somehow), and tries to run every conceivable e2e test once it does literally anything.
like yesterday it ran for like an hour to build a fairly basic frontend...
i like luna and sol but it feels bad lately
Have you tried Grok 4.6, if you're focused on token budgets? In a league of it's own for tokens/intelligence.
Not for enterprise. Can't trust the company behind it with my data.
SuperGrok quota is garbage for anything coding. I burn through my quota in a few hours with very mild use.
SuperGrok Plus is slightly better but doesn’t last me more than a few days. Even Claude Max feels leagues more generous in usage…
I haven’t tried SuperGrok Heavy because it’s too expensive
All of the subscription AI platforms are trimming down quotas across the board to push users into higher tiers. Whatever they can do. Local inference needs to meet pricing sooner