My bias here is that I’m in the market to spend on some shiny new hardware. I like DIY tech, and I only have a really old laptop for it to replace.

With the mounting cost of AI subs, and the variety of my own personal use cases, can I find a way to convince myself to buy the M5 Ultra Mac Studio?

These are all my own calculations and ideas, and thus may betray a lack of knowledge. If you see bad assumptions that would change the outcomes, I’d love to hear from you!

Local models have improved a lot both in objective benchmark measures and subjective experience (as reported by others, I’ve only used local transcription in anger). They are a long way behind what’s available at the frontier, especially on realistic hardware setups. So, rather than treat them as a direct replacement for the full stack of solutions provided by one of the big companies, I’ll aim to replace a specific usage type with local models.

My usage falls into three categories:

1. Software engineering - Background continuous work
2. Personal assistant - The main orchestrating agent
3. Google replacement - Standard chatbot stuff

For engineering, there’s also interactive use where I’m actively involved in the work with something like 10 to 100 agents/sub agents running. There’s no local hardware, that I’m aware of, that can run this number of sessions with reasonable context size. A hybrid approach, using local for some tasks, could work but I think the higher potential use case is non-interactive background work. This seems ideal as it can just soak up any unused capacity.

Today’s solutions

1. Software Engineering - Almost exclusively interacting through Codex Desktop/iOS. An OpenAI Pro sub, plus Claude Max for access to Fable.
2. Assistant - OpenClaw on a Mac mini using 5.6 Sol. (Keen to go all Codex but my Mac mini is 16GB and Codex often gets OS paused for using too much memory. My other device is an Intel Mac, which doesn’t support computer use. So I’m stuck… until I get the Studio.) ;)
3. Google replacement - ChatGPT desktop/iOS app, largely voice

Candidate models

• Qwen3.8‑27B - A 30B class model tuned for code. This seems to be the size local models are converging on. I could run multiple threads with reasonable context.
• DeepSeek V4 Flash 0731 (MoE 284/13) - As it’s a single orchestrator, I want the largest model I can fit on the machine. Tokens per second (TPS) matter as it’s often interactive. DeepSeek compressed attention means I can get up to 1M context fairly comfortably.
• GLM‑5.3‑Flash (320B/18) - The highest intelligence model that can fit on device. Context less important as most of these conversations are short.

I’ve taken my usage from the last month to get a ballpark figure for how much capacity is needed:

• Software engineering - This is probably the most variable. But it’s fairly easy to saturate both subscriptions if I’m actively working on something. They are personal projects so there’s flexibility to use capacity or not, which means on the order of 10B tokens (blended) per month.
• Assistant - In August: 98.88M in, 1.96B cached, 2.16M out. I comfortably manage this within my OpenAI sub, ramping down dev work if I get close to my weekly limits.
• Google replacement - I’ve requested my data from OpenAI and will update the guess with real data. Estimate 500 chats/month, 40k in / 3k output tokens on avg.

Requirements set.

Is there a cost case?

Frontier companies make huge margins, especially on the API pricing. Can I run things locally and avoid paying all that vig? It seems that for my use cases, probably not. Open models, of the size I could run locally, are priced quite competitively and subscription pricing for frontier models is hard to pass up.

To justify the hardware spend I’ll calculate a theoretical “time to payoff”. Having the hardware is a reward in itself for me and at some point it stops losing value, so a full payoff might not be a fair bar. Your weightings may differ though, and I don’t know what the future holds, so full payoff is what I’ll use. I’m also assuming fixed prices for simplicity.

At point of writing, the best prices I could find on OpenRouter or Vercel

Hardware costs

For the coding use case, I assume 24/7 background coding. For the other two, I use the estimated token usage, since having more “free” tokens wouldn’t change how much I use them. For the assistant, the hosted equivalent is the £20 subscription rather than API pricing, since that’s what I’d actually pay.

Potentially I could make the case for a Spark paying itself back under heavy consistent usage. But realistically it should be compared to something like 5.6 Luna, which benches slightly better across the board and gives a comparable number of tokens on the £20 sub. This is all before I factor in the “cost” to manage the local models: upgrades, security fixes, debugging, downtime etc. Some of this would be fun and kind of the point of it all. But I think in the long term, this should probably be considered a downside for a system I depend on.

Another assumption is that these smaller models are what I would use given the choice. For me, at the moment, this isn’t the case. Certainly for the chat and orchestration use case I could probably get away with the £20 a month OpenAI sub and be using the top-tier model, which seems like a no-brainer.

For engineering work there’s more of a case, especially as model routers get better. For low-intensity interactive or background work I could save a little bit vs using cloud versions of the same models. But subscriptions again blow these figures out of the water.

With subscriptions the tradeoff in quality/reliability to use a smaller, slower model isn’t worth it. The additional cost of running a better model for the task is usually trivial vs the cost of something going wrong and needing to fix it (and it’s effectively 0 as I would have the subscription to orchestrate anyway). But if I had to pay API pricing for everything, it would be a much more involved decision.

So cost doesn’t appear to be a justification for me to splurge. What other reasons can I come up with?

Why use local models in the first place?

Here are the other motivations for using local models, set against my use cases.

There is some value in having more control but I think this is more often outweighed by having access to better and improving models. The privacy angle isn’t something I weigh highly in my own usage, certainly not enough to dramatically change my usage patterns. And, although it’s fun to have offline models, it’s pretty hard to be offline these days, even on a plane.

Personally, I don’t see local models replacing my current use cases at the moment. The convenience and quality available on the frontier, especially at subscription pricing, is too good to give up, and I can get the local model learning experience with smaller models and no stakes, rather than depending on them.

Perhaps the price of memory will keep rising and the whole thing turns out to be a straight investment with returns to be had!