Rent or Buy? What Running the Big Three Actually Costs — Tokens, Hours, or Silicon
Status board (checked 2026-11-04)
Layer 1 — per token (OpenRouter aggregator, checked Nov 3; per 1M tokens in / out):
| Model | Aggregator | Official direct |
|---|---|---|
| GLM-5.3-Flash | $0.0313 / $0.0846 | $0.15 / $0.50 (z.ai) |
| GLM-5.3 (flagship) | $0.08 / $3.60 | $1.40 / $4.40 (z.ai) |
| DeepSeek-V4.1-Flash | $0.045 / $0.50 | check official page |
| Qwen3.8-Flash | $0.15 / $0.47 | check official page |
| MiniMax-M3.1-Flash | $0.30 / $1.20 (M3 listing) | $0.15 / $0.60 flat (MiniMax, Oct 19) |
| Kimi K3 | $0.20 / $15.00 | check official page |
Layer 2 — per hour (Chinese spot-rental reports, late Oct): H100 $1.89–2.75 · L40S $1.33–2.04 · A100 $1.10–1.79 · RTX 4090 ≈ ¥2.3/hr spot, ¥2.2–2.88 on science-platform lists.
Layer 3 — own silicon: 2×5090 rig ≈ $6,600 at Nov street prices (see the GPU board) — and the RAM bill rides the same squeeze.
What it is
Three different ways to pay for the same tokens, and they answer different questions:
- Aggregator/API — zero hardware, zero fixed cost; the Flash class is priced like a utility.
- Hourly rental — someone else’s 4090/H100 by the hour; right for bursts, wrong for 24/7.
- Own rig — fixed cost, no per-token bill, weights and data stay in your box (why that matters).
Getting started (3 steps)
- Estimate your tokens. At the aggregator’s GLM-5.3-Flash rate, a heavy 10M-token month costs ≈ $1 — no rig on earth beats that price; the rig competes on control, not cost.
- Do the break-even honestly. Formula:
rig price ÷ (hourly rate × hours/month) = break-even months. Example: $6,600 ÷ (≈$1.6/hr × 24×30) ≈ 6 months at 24/7; at 8 h/day it’s ~16 months. Below ~50% duty cycle, renting wins. - Trial on Layer 2 before buying Layer 3 — rent the exact card class for a day, run your real workload, then decide. Spot instances can be preempted; plan around it.
Best use
| You are… | Cheapest honest path |
|---|---|
| Testing prompts / light daily use | Aggregator Flash-class — pennies, no hardware, no commitment |
| Running agent bursts a few hours/week | Hourly rental of the card class you’d have bought |
| Serving 24/7, privacy-sensitive, or fine-tuning daily | Own rig — and note aggregator flagship prices ($3.60 out for GLM-5.3) make heavy use add up fast |
Honest weaknesses
- Aggregator prices float daily; the table above is a Nov-3 snapshot, and the official pages always win.
- Spot instances get preempted; long jobs need checkpointing or on-demand pricing (2–3× spot).
- Rented boxes start empty: downloading 250 GB of weights and lookup tables before the first token is on your clock.
- Domestic platforms bill in ¥ with their own signup quirks; overseas aggregators bill in $.
FAQ
Why is the aggregator cheaper than z.ai direct for the same model?
Competition among providers on the aggregator drives Flash down to ~1/5 of list — the same gap shows on the flagship ($0.08/$3.60 vs $1.40/$4.40). The lab’s own endpoint is the ceiling, not the floor.
Isn’t self-hosting always cheaper “in the long run”?
Only above a duty cycle you probably don’t have. At Flash prices the break-even is years; at flagship prices it’s months. Compute both before believing either.
What about the free tier?
The tools pages track each assistant’s free chat tier. For API-style use, new accounts get trial credits (e.g. z.ai’s promo tokens) — but the durable free lunch in 2026 is the open weights themselves.
📌 Summarised from official docs and hands-on samples. Product features and pricing change often — official site wins.
← Newer: How Much RAM — and It Costs What Now? The 2026 Memory Squeeze, Checked November Older: What GPU Do You Need to Run the Big Three at Home? VRAM Tiers, Checked November 2026 →Related: Which Software Runs the Big Three? Strata, Maya, and the GGUF World, Compared
What Quantization Costs You: the IQ1→Q8 Ladder With Measured Numbers
What GPU Do You Need to Run the Big Three at Home? VRAM Tiers, Checked November 2026