AI Cost Calculator

Rule of thumb: if you spend more than ~$200/month on API tokens consistently for the model class you want, local hardware breaks even within 12-18 months. At $50/month API spend, you'll never recover a $4,000 Mac Studio purchase from API savings alone — the math only works if you also value privacy, offline capability, or unlimited fine-tuning. The calculator above runs the exact numbers based on your token volume, output ratio, and electricity rate.

Yes — pulled from the public pricing pages of OpenAI, Anthropic, Google, DeepSeek, Mistral, and Together AI as of May 2026. Update window: within 7 days of any major price change. Note that most providers have rate-discount tiers, batch-API discounts (50% off), prompt-caching discounts (90% off cached tokens), and enterprise-volume discounts that aren't reflected here. For high-volume production use, your effective API cost will be 30-70% lower than the calculator shows.

Hardware purchases are capital expenditure that should be spread over its useful life for fair comparison. We use 3-year straight-line amortization (industry standard for GPUs). A $4,000 Mac Studio costs $111/month over 3 years. After 3 years, the marginal cost drops to just electricity. If you keep the hardware longer, your effective monthly cost decreases further; if you upgrade sooner, it increases.

Local hardware costs are dominated by hardware amortization, not electricity, unless you run the GPU at high utilization 24/7. Utilization 20% means the GPU is at full power for ~5 hours/day on average — typical for a developer's personal machine. For production serving with sustained traffic, set utilization to 80-100% to model true continuous load. This is the single biggest lever in the local cost calculation: a 20% → 100% utilization change can multiply electricity cost 5x.

No — it's pure compute cost. Operating local infrastructure requires engineer hours: setup (4-40 hours), monitoring (2-10 hrs/month), upgrades (4-8 hrs/quarter), debugging (variable). At $200/hr fully-loaded engineer cost, the ops overhead can match or exceed compute cost for small teams. APIs trade higher per-token cost for zero ops overhead. For solo devs and SMBs, this is the dominant trade-off — not the per-token math. For 50+ engineer teams, ops cost is a rounding error and per-token cost dominates.

Hybrid is what most production teams actually run. Pattern: route 70-90% of queries to a self-hosted model (covers easy / repetitive / long-tail traffic), route the hardest 10-30% to a frontier API (Claude Opus 4.7 / GPT-5.5 Pro / Gemini 3.1 Pro). Total cost: roughly 30% of pure-API for the same quality envelope. Tools like LiteLLM, Portkey, and OpenRouter make hybrid routing trivial. Use this calculator for the local-portion math; estimate API spend on the remaining frontier-tier traffic separately.

For personal / single-user inference, the Mac Studio is appropriate even at 32B-70B model sizes. For serving multiple concurrent users, you need an H100 (or 2x H100) running vLLM with continuous batching — single-stream Mac throughput becomes a bottleneck once you have more than 1-2 simultaneous users. The calculator assumes single-user inference patterns. For production multi-user serving, switch the hardware option to "2x H100 cluster" and adjust utilization to your real production load.

Together AI (and other inference providers like Fireworks, Replicate, Anyscale) host the same open-weight models you would self-host, but they amortize their H100/H200 cluster across thousands of customers. At low volume (< 50M tokens/month), their per-token economics beat yours. At high volume (> 500M tokens/month), self-hosting wins because you pay once for hardware vs ongoing per-token markup. Use Together / Fireworks for prototyping and low-volume production; switch to self-host once monthly spend exceeds ~$1,000/month.