⚡ Cost Calculator · tuned for Qwen 3.8 27B

Estimate electricity costs for running your local LLM server. Presets and the GPT 5.6 Luna comparison are calibrated for Qwen 3.8 27B — the closest local match to GPT 5.6 Luna in intelligence.

Parameters

$/kWh
W
tok/s
tok/s
tok
%
tok
req

Results

Cost / Hour
$0.00
continuous operation
Cost / 1M Input · no cache
$0.00
0% cached
Cost / 1M Input · all cached
$0.00
100% cached (100× faster)
Cost / 1M Input · realistic
$0.00
per 1M input tokens (with cache)
Cost / 1M Output
$0.00
per 1,000,000 output tokens
Per Request
$0.00
— s
Daily Cost
$0.00
— requests
Monthly Cost
$0.00
30-day estimate
Input tok / kWh
—
with cache
Output tok / kWh
—
decode only

🎯 Realistic Scenario

PC working 90% of the day (21.6 h). Average request: 60k input tokens (50k cached), 2k output tokens. Uses your current power / energy / speed settings above.
Daily (90% uptime)
$0.00
21.6 h active
Monthly (90% uptime)
$0.00
30-day estimate
Per Request
$0.00
— s
Req / Day (sustained)
—
max requests in 21.6 h
Input tok / kWh
—
83% cached (50k/60k)
Output tok / kWh
—
decode only

📊 GPT 5.6 Luna Comparison

Self-hosted Qwen 3.8 27B profiles vs GPT 5.6 Luna API ($0.25/1M input, $0.02/1M cached, $1.20/1M output). Qwen 3.8 27B is the closest local match to GPT 5.6 Luna in intelligence, so this is a fair like-for-like comparison. Scenario: 60k input (50k cached) + 2k output per request, 90% uptime (21.6 h/day).
Profile Power Speed Per Request Monthly (self) Monthly (GPT) Savings

⚡ Power Efficiency

Profile Input tok/kWh Output tok/kWh
How it works
cost_per_hour = (power_W / 1000) × energy_cost_$
uncached_tok = prompt_tokens × (1 − cache_pct / 100)
cached_tok = prompt_tokens × (cache_pct / 100)
prompt_time = uncached_tok / prompt_speed + cached_tok / (prompt_speed × 100)
request_time = prompt_time + output_tokens / decode_speed
cost_per_request = request_time / 3600 × (power_W / 1000) × energy_cost_$
cost_per_1M_input = (1M × (1−c) + 1M × c/100) / prompt_speed / 3600 × (power_W/1000) × energy
daily_cost = cost_per_request × requests_per_day
monthly_cost = daily_cost × 30
scenario_daily = (power_W / 1000) × energy_cost_$ × 24 × 0.9

Cached tokens are processed at 1/100 the time of uncached tokens (KV-cache hit). The realistic scenario assumes the PC draws full power for 90% of the day (21.6 h) — a worst-case continuous-operation estimate.