The True Cost of Running AI Locally — £0.08/M Tokens vs £24/M Tokens

I calculated the real cost of local+OAuth AI inference including hardware amortisation and electricity. The result: 310x cheaper than Claude Opus, 62x cheaper than Sonnet, and even 3x cheaper than GPT-4o mini.

2026-04-17 · 5 min read

215 Million Tokens for £16.64/Week

Everyone talks about the cost of AI APIs. Nobody talks about the cost of running it yourself — including the hardware you already bought.

I tracked every token for a full week across my stack: GLM-5.1 and GPT-5.3 Codex on flat-rate OAuth subscriptions (£4.62/wk each), Qwen3.5 9B running locally on a consumer RTX 5070 Ti, plus free-tier cloud models for light tasks.

Here’s the headline:

That’s right. Including everything — the GPU, the RAM, the electricity, the subscriptions — I am processing 215 million tokens per week at eight pence per million tokens.

The Stack, Itemised

*Local inference electricity is included in the £1.15 figure. The model itself is free.

How I Got Hardware Amortisation

I am not going to pretend hardware is free — that’s the trick most "local AI is cheaper" articles pull. Here’s my maths:

• AI-relevant hardware: RTX 5070 Ti (about £750), 64GB DDR5 (about £200), Core Ultra 7 265KF (about £350) = about £1,300
• Useful life: 4 years (208 weeks)
• Weekly amortisation: £6.25/week

Yes, you could argue the PC gets used for other things too. And you’d be right — if your GPU is also your gaming rig, cut that in half. But even at full price, it’s a rounding error compared to per-token API costs at the workload I’m running.

The Comparison That Matters

Here’s where it gets fun. I took my actual token usage and calculated what it would cost at published per-token pricing:

Three hundred and ten times cheaper than Opus. Sixty-two times cheaper than Sonnet. Even against the cheapest per-token API — GPT-4o mini at £0.26/M — I am still 3x better off.

And before you say "but GPT-4o mini isn’t as good" — you’re right, it isn’t. GLM-5.1 and GPT-5.3 Codex are genuinely powerful models. This isn’t a toy comparison.

The Input/Output Imbalance

My token mix is heavily input-biased — 99.4% input, 0.6% output. This is typical for agent workloads: tool results, web pages, and file contents dominate the context window, while the model’s responses are relatively terse.

This is important because API pricing heavily penalises output tokens. Claude Opus charges £6/M input but £30/M output. When your workload is 99% input, the "cheap input" headline rate is misleading — you’re still paying through the nose because the per-token model treats your 200M input tokens as a revenue opportunity.

Flat-rate subscriptions flip this: whether your ratio is 1:1 or 220:1, the price stays at £4.62/week.

Why "Free Local" Isn’t Actually Free

Qwen3.5:9b runs locally and costs £0 in API fees. But I included its electricity in my calculation because honesty matters. At ~350W system draw and roughly 6.5 GPU-hours over the week, local inference adds about £1.15/week to the electricity bill.

That’s trivial — but it’s not zero. And if you’re running agents 24/7, that number climbs fast. A constantly-occupied GPU at 350W over a full week is 58.8 kWh, which at UK rates is about £17/week — more than the subscriptions.

The lesson: local inference is "free" until you saturate the GPU. Then electricity becomes your new per-token cost.

What You Actually Need

The Fine Print

• My effective rate fluctuates. Light weeks = higher per-token cost. Heavy weeks = lower. At 215M tokens/week, £0.08/M is my current rate. It’ll settle further as I run more agents concurrently.
• Flat-rate plans have rate limits. You’re not getting unlimited throughput — you’re getting predictable cost. If you need 10 concurrent sessions hammering Opus, OAuth won’t save you.
• Hardware costs are front-loaded. You pay £1,300 on day one. The amortisation is comforting on paper, but you still paid it already.
• Local models have quality ceilings. Qwen3.5:9b handles routing and simple tasks well. It doesn’t replace GLM-5.1 for complex reasoning. That’s why I have both.

The Bottom Line

The "true cost of running locally" isn’t just the electricity. It’s subscriptions + hardware + electricity. But even accounting for all of it, the maths is brutal for per-token APIs.

£16.64/week for 215M tokens. That’s £0.08 per million tokens including everything.

The same volume on Claude Opus would cost £5,167/week. That’s not a typo. That’s three hundred and ten times more expensive.

The Cost Comparison — At a Glance

View full-size infographic

Run locally. Run smart. Run the numbers.

Found this useful? 👉 Follow @Raf_VRS for more Hard Interference field notes. 👉 Support the work: ko-fi.com/rafvrs