Cost modeling and ROI analysis for specialized LLM development. Use when deciding whether to train a custom model, estimating total cost, or calculating break-even vs frontier APIs. Covers training costs, inference costs, and time-to-ROI projections.
한국어 요약은 준비 중입니다. 원문 설명을 표시하고 있습니다.
설치하기 전에 에이전트에게 전달되는 지시 내용을 확인할 수 있습니다.
Use this skill as the first step of any model-specialization assessment to determine:
Important: Use websearch to get current pricing — these change frequently. Reference tiers as of early 2026:
| Model | Input ($/M) | Output ($/M) | Notes |
|---|---|---|---|
| Claude Opus 4.6 | $5.00 | $25.00 | Highest capability |
| GPT-5.2 Pro | $21.00 | $168.00 | Reasoning-heavy |
Savings potential: Highest. A specialized 8B model costs ~$0.05-0.15/M tokens to serve = 100-500x cheaper.
| Model | Input ($/M) | Output ($/M) | Notes |
|---|---|---|---|
| GPT-5.2 Chat | $1.75 | $14.00 | Most popular |
| Claude Sonnet 4.5 | $3.00 | $15.00 | Good balance |
| Qwen3 Max | $1.20 | $6.00 | Cost-effective |
Savings potential: Strong at volume. Break-even is achievable with moderate traffic.
| Model | Input ($/M) | Output ($/M) | Notes |
|---|---|---|---|
| Gemini 3 Flash | $0.50 | $3.00 | Fast, cheap |
| GLM 4.7 | $0.40 | $1.50 | Very cost-effective |
| Mistral Small | $0.10 | $0.30 | Cheapest commercial |
Savings potential: Harder to beat on cost alone. Specialize for quality, not cost.
Many open-weight models available on OpenRouter at $0. If you only need basic capability, self-hosting these is essentially free beyond GPU cost.
Cheapest option for supported models. No GPU management.
| Model Size | Method | Cost/1K examples | Duration |
|---|---|---|---|
| 8B | LoRA | ~$2-5 | 15-30 min |
| 14B | LoRA | ~$5-10 | 30-60 min |
| 32B | LoRA | ~$10-25 | 1-3 hours |
| 70B | LoRA | ~$25-60 | 3-8 hours |
Load tinker-training-cost skill for exact per-token rates.
| GPU | Cost/Hour | Best For |
|---|---|---|
| A100-40GB | ~$1.10 | 7-14B LoRA, 7B full |
| A100-80GB | ~$1.60 | 14-32B LoRA, 14B full |
| H100-80GB | ~$3.25 | 32-70B LoRA, large batches |
Typical training runs:
| GPU | Cost/Hour | Best For |
|---|---|---|
| A100-80GB | ~$1.50 | Standard training |
| H200-141GB | ~$3.00 | Large models, fast training |
| B200-192GB | ~$4.50 | Cutting-edge, maximum throughput |
Best for: Multi-day training, large-scale RL, dedicated inference.
The key calculation: what does it cost to serve the specialized model vs the frontier API?
def inference_cost_per_million_tokens(gpu_cost_per_hour, tokens_per_second):
"""Calculate cost per million tokens for self-hosted inference.
Args:
gpu_cost_per_hour: GPU rental cost (e.g., $3.25 for H100)
tokens_per_second: Model throughput (varies by model size and batch)
"""
tokens_per_hour = tokens_per_second * 3600
cost_per_token = gpu_cost_per_hour / tokens_per_hour
cost_per_million = cost_per_token * 1_000_000
return cost_per_million
| Model Size | GPU | Tokens/sec (batch) | Cost/M tokens |
|---|---|---|---|
| 8B | A100-80GB | ~3,000-5,000 | $0.09-0.15 |
| 8B | H100-80GB | ~5,000-8,000 | $0.11-0.18 |
| 14B | A100-80GB | ~1,500-3,000 | $0.15-0.30 |
| 14B | H100-80GB | ~3,000-5,000 | $0.18-0.30 |
| 32B | H100-80GB | ~800-1,500 | $0.60-1.10 |
| 70B | 2xH100 | ~500-1,000 | $1.80-3.60 |
Key insight: A specialized 8B model on Modal/TensorPool costs $0.05-0.15/M tokens vs $5-25/M for premium frontier = 50-500x cheaper per request.
def break_even_analysis(
monthly_api_spend: float,
training_cost: float,
monthly_inference_cost: float,
setup_hours: float = 40,
hourly_rate: float = 100,
):
"""Calculate break-even timeline for model specialization.
Args:
monthly_api_spend: Current monthly frontier API cost
training_cost: One-time training cost (GPU + data prep)
monthly_inference_cost: Monthly cost to serve specialized model
setup_hours: Engineering time to set up pipeline
hourly_rate: Engineering cost per hour
"""
engineering_cost = setup_hours * hourly_rate
total_upfront = training_cost + engineering_cost
monthly_savings = monthly_api_spend - monthly_inference_cost
if monthly_savings <= 0:
return {
"viable": False,
"reason": "Specialized model costs more to serve than frontier API",
"monthly_savings": monthly_savings,
}
months_to_break_even = total_upfront / monthly_savings
return {
"viable": True,
"total_upfront_cost": total_upfront,
"training_cost": training_cost,
"engineering_cost": engineering_cost,
"monthly_api_spend": monthly_api_spend,
"monthly_inference_cost": monthly_inference_cost,
"monthly_savings": monthly_savings,
"months_to_break_even": round(months_to_break_even, 1),
"annual_savings": monthly_savings * 12 - total_upfront,
"year_2_savings": monthly_savings * 12, # Fully amortized
}
# Example: Replace GPT-4o for a specific task
result = break_even_analysis(
monthly_api_spend=5000, # $5K/month on GPT-4o
training_cost=500, # $500 for LoRA fine-tune + data prep
monthly_inference_cost=200, # $200/month on Modal (8B model)
setup_hours=40, # 1 week of engineering
hourly_rate=100, # $100/hr engineer
)
# Result: Break-even in ~1.1 months, $52K annual savings after year 1
| Month | Frontier Cost | Specialized Cost | Cumulative Savings | Notes |
|---|---|---|---|---|
| 0 | — | $4,500 upfront | -$4,500 | Training + engineering |
| 1 | $5,000 | $200 | $300 | First month live |
| 2 | $5,000 | $200 | $5,100 | Break-even |
| 3 | $5,000 | $200 | $9,900 | |
| 6 | $5,000 | $200 | $24,300 | |
| 12 | $5,000 | $300 | $52,200 | Includes retrain cost |
Route easy requests to specialized model, hard ones to frontier:
Ask the user these to determine viability:
What's your monthly frontier API spend?
How constrained is the task?
What production data do you have?
What's your quality bar?
아직 리뷰가 없습니다. 사용해 본 소감을 남겨 주세요.
로그인 없이 쓸 수 있습니다. 사용 후기, 질문, 작동하지 않은 환경 등을 편하게 남겨 주세요.
아직 댓글이 없습니다. 첫 댓글을 남겨 보세요.
Simplest distributed training API. 4 lines to add distributed support to any PyTorch script. Unified API for DeepSpeed/FSDP/Megatron/DDP. Automatic device placement, mixed precision (FP16/BF16/FP8). Interactive config, single launch command. HuggingFace ecosystem standard.
한국어 요약은 준비 중입니다. 원문 설명을 표시하고 있습니다.
ADMET property prediction for drug candidates. Full pharmacokinetic panel (Caco-2, PPB, clearance, CYP), toxicity (hERG, AMES, DILI), drug-likeness (Lipinski, QED), using RDKit descriptors and TDC models.
한국어 요약은 준비 중입니다. 원문 설명을 표시하고 있습니다.
Interpretable ADMET analysis with mechanistic reasoning. Maps liabilities to structural causes and biological pathways. Based on CoTox (Park 2025) and DrugR (Liu 2026).
한국어 요약은 준비 중입니다. 원문 설명을 표시하고 있습니다.
This skill should be used for time series machine learning tasks including classification, regression, clustering, forecasting, anomaly detection, segmentation, and similarity search. Use when working with temporal data, sequential patterns, or time-indexed observations requiring specialized algorithms beyond standard ML approaches. Particularly suited for univariate and multivariate time series analysis with scikit-learn compatible APIs.
한국어 요약은 준비 중입니다. 원문 설명을 표시하고 있습니다.
Infer gene regulatory networks (GRNs) from gene expression data using scalable algorithms (GRNBoost2, GENIE3). Use when analyzing transcriptomics data (bulk RNA-seq, single-cell RNA-seq) to identify transcription factor-target gene relationships and regulatory interactions. Supports distributed computation for large-scale datasets.
한국어 요약은 준비 중입니다. 원문 설명을 표시하고 있습니다.
Comprehensive Python library for astronomy and astrophysics. This skill should be used when working with astronomical data including celestial coordinates, physical units, FITS files, cosmological calculations, time systems, tables, world coordinate systems (WCS), and astronomical data analysis. Use when tasks involve coordinate transformations, unit conversions, FITS file manipulation, cosmological distance calculations, time scale conversions, or astronomical data processing.
한국어 요약은 준비 중입니다. 원문 설명을 표시하고 있습니다.