Cloud vs. Rig TCO
Plot cumulative 36-month API spending against hourly RunPod/H100 rentals and owned RTX/Mac hardware with electricity and resale value.
See why 12-turn tool-calling agent loops multiply context length and how 90% prompt caching rescues your margin.
Multi-turn history expands average input context to 15,450 tokens/req. Output accounts for 57% of total spend.
| Model | Delta | |||
|---|---|---|---|---|
| Claude Haiku 5.5 (High-Volume)Anthropic | $0.10 | $0.50 | $50.32 | Save $903 |
| GPT-4.1 nano (Ultra-Low Cost)OpenAI | $0.10 | $0.40 | $52.78 | Save $901 |
| DeepSeek V4.1 Flash (Off-Peak)DeepSeek | $0.15 | $0.60 | $60.91 | Save $893 |
| GPT-4o mini (Legacy Budget)OpenAI | $0.15 | $0.60 | $99 | Save $854 |
| Gemini 3.5 Flash-Lite (Budget Edge)Google | $0.30 | $2.50 | $205.74 | Save $748 |
| GPT-4.1 mini (Balanced Fast)OpenAI | $0.40 | $1.60 | $211.1 | Save $742 |
| DeepSeek V4 Pro (Off-Peak Reasoning)DeepSeek | $0.66 | $1.98 | $236.51 | Save $717 |
| Gemini 3.8 Flash (Intro Paid Tier)Google | $0.75 | $3.75 | $377.39 | Save $576 |
| Llama 3.3 / 4 70B-Class (Hosted API)Open-Weight Host | $0.59 | $0.79 | $459.46 | Save $494 |
| Gemini 4 HydrogenGoogle | $1.50 | $7.50 | $715.11 | Save $238 |
| o3-mini (Reasoning)OpenAI | $1.10 | $4.40 | $726.02 | Save $227 |
| GPT-5 (Flagship Reasoning)OpenAI | $1.25 | $10.00 | $834.44 | Save $119 |
| Claude Sonnet 5.5 (Coding & Agents)Anthropic | $2.00 | $10.00 | $953.47 | Active Baseline |
| GPT-6.2 SolOpenAI | $2.00 | $10.00 | $953.47 | +$0 |
| Gemini 4 CarbonGoogle | $2.00 | $10.00 | $953.47 | +$0 |
| GPT-4.1 (1M Long-Context)OpenAI | $2.00 | $8.00 | $1,055.52 | +$102 |
| Gemini 3.1 Pro (Long-Context)Google | $2.00 | $12.00 | $1,115.95 | +$162 |
| Gemini 4 NeonGoogle | $3.00 | $15.00 | $1,430.21 | +$477 |
| GPT-4o (Legacy Omni)OpenAI | $2.50 | $10.00 | $1,650.05 | +$697 |
| Claude Opus 5.5 (Agentic Frontier)Anthropic | $4.00 | $20.00 | $1,906.95 | +$953 |
| Claude Fable 5.2Anthropic | $8.00 | $40.00 | $3,708.09 | +$2,755 |
| GPT-6.2 AstraOpenAI | $8.00 | $40.00 | $3,813.89 | +$2,860 |
Without Anthropic 90% prompt caching, multi-turn coding agents spike dramatically. Pre-populated at 1,500 daily requests with 12,000 input tokens, 1,200 output tokens, 75% prompt caching, and 6 turns/session.
Primary Technical & US Grid Benchmarks
Plot cumulative 36-month API spending against hourly RunPod/H100 rentals and owned RTX/Mac hardware with electricity and resale value.
Count live prompt tokens across English prose and JSON/code, size RAG document chunks, and compute float32 vs. int8 HNSW vector RAM.
Test Llama 70B, Qwen 32B, or DeepSeek R1 across RTX 3090/4090/5090 & Apple Silicon Macs with GQA KV-cache math and tokens/sec speed.