I Used 1.5 Billion Tokens — Should I Still Buy a Mac Mini?

If you use AI agents long enough, you eventually face this question.

Is it better to keep paying for APIs? Or buy a machine like a Mac Mini and run local AI instead?

I've been wrestling with this too. Not because I just want new hardware. If I could reduce what I spend on paid APIs, buying local AI equipment would be worth considering.

So instead of vague speculation, I checked my actual usage records.

1. I Use Two Paid APIs

I currently have $20 charged to each of DeepSeek and MiMo, for a total of $40.

The charge date was around July. I'm still using both services, but there's balance left.

DeepSeek Usage

Here are the figures from the DeepSeek dashboard.

About 80% of the charged amount has been used. There's still balance left, so I can keep using it without recharging.

The dashboard also shows a separate Cost $8.37 item. However, the aggregation period for this item is hard to confirm from the screen alone, so I did not add it to the total cost.

MiMo Usage

MiMo's cumulative spend is $13.31.

A significant portion of MiMo's input tokens are counted as cache hits.

Across all token records, cache-hit input accounts for about 90.4%. Among input tokens, about 90.9% are cache hits.

This matters because if frequently used context or prompts are served from cache, different rates may apply. Actual cost depends on the model and billing conditions, but high token volume does not necessarily mean high cost.

2. Combining Both APIs' Actual Usage

I added up the figures from both services.

Simply summing the tokens shown by both services gives about 1.588 billion.

Of course, the token counting methods of two different services may not be exactly the same. The input/output composition also differs, and billing varies based on cache hits, so token count alone cannot determine which service is more efficient.

Still, looking at the actual volume and cost together reveals something interesting.

I processed a very large number of tokens, yet total spend is just under $30.

Charged around July, used for about 3 months — that's a simple monthly average of about $9.79. The exact monthly figure may vary based on the actual charge date and billing period, but based on the usage pattern so far, API cost is not a heavy burden.

3. Then Is Buying a Mac Mini Necessary?

I was considering a Mac Mini to save on paid API costs.

The idea was to run tasks that local models can handle on my own PC, and use paid APIs only for difficult reasoning or complex code review, reducing costs.

But calculating based on actual usage records changed my thinking a bit.

Assume the current monthly average API spend is about $9.79.

• Cutting API costs by 50% saves about $4.90 per month.
• Cutting API costs by 80% saves about $7.83 per month.
• Eliminating API costs entirely saves about $9.79 per month.

Of course, this calculation assumes the recent usage pattern continues. If agent work increases and API consumption surges, the results could differ.

But based on current spending, buying new hardware just to cut paid API costs doesn't look very economical.

Moreover, local models cannot completely replace paid APIs. You may still want to use stronger models for complex reasoning or tricky code reviews.

Ultimately, to recover the Mac Mini's price, you'd need to figure out how much API cost it can actually replace.

4. Then Is Local AI Unnecessary?

That's a different story.

I already use local models on a PC with an RTX 3070. I handle frequently repeated tasks with local models and pass difficult tasks to APIs like DeepSeek.

Adding a Mac Mini could have meaning beyond just reducing API costs.

It could be for testing larger models that my current PC can't run, running AI agents on separate hardware full-time, or studying local AI systems themselves.

For these purposes, buying equipment isn't just about cost savings — it's investing in a new experimental environment.

But looking at my actual usage alone, costs haven't grown to the point where the lack of extra hardware makes paid API costs unbearable.

So there's no need to rush the purchase.

Conclusion — Look at Usage Patterns Before Equipment

When buying AI equipment, it's easy to focus on performance, memory, and local model capabilities first. I'm also interested in hardware that can run larger models.

But this time, I looked at actual API usage records first.

I charged $20 each to DeepSeek and MiMo, and the total cumulative spend shown so far is $29.37. Summing the token records gives over 1.5 billion. Yet my monthly average spending is around $10.

At this level, buying a Mac Mini solely for token cost savings doesn't seem necessary.

If agent work increases and costs surge, or if a concrete need for larger local models emerges, I can recalculate then.

The question isn't whether to buy hardware first, but what work can be handled locally and how much cost it can replace.

My conclusion is simple.

Before buying new hardware, let's make better use of the current local models and paid APIs. The Mac Mini can wait.