o4-mini-2025-04-16· OpenAI
LMArena Human Preference 140K
what would be the cost
- User
what would be the cost
- Assistant
Here’s a rough breakdown of the compute‐only cost if you build on top of DeepSeek-R1, using H100 spot/ondemand rates:
Assumed H100 rate Spot – 2.00/hr
Phase H100-hrs Cost @ $1/hr Cost @ $2/hr Domain Adaptation / Fine-tune 1 000–20 000 $1 000–20 000 $2 000–40 000 Sim-to-Real Transfer 10 000–100 000 $10 000–100 000 $20 000–200 000 RL / Policy Optimization 10 000–100 000 $10 000–100 000 $20 000–200 000 Real-World Fine-tuning 1 000–10 000 $1 000–10 000 $2 000–20 000 Total 22 000–230 000 $22 000–230 000 $44 000–460 000 Key takeaways:
- At 44K–$460K of GPU‐only costs.
- Using spot instances (≈ $1/hr) or negotiated enterprise pricing can roughly halve this.
Keep in mind this excludes:
• Data storage, preprocessing, and networking fees
• Engineering time (data collection, annotation, integration)
• On-prem hardware depreciation (if you buy your own H100s)But purely for cloud GPU time, plan on tens to a few-hundred thousand dollars.
Other models, same or similar prompt
Responses were produced in different, uncontrolled contexts (system prompts, settings, prior turns). This is not a benchmark.