OpenAI vs Self-Hosting Cost Calculator
Should you keep paying OpenAI per token, or rent a GPU and run an open-weights model instead? This page is the TokenForge calculator preset for exactly that question: GPT-5.6 Terra against a rented H100. Adjust the volume, tier and server to match your setup — prices refresh daily.
OpenAI vs self-hosting tool
At 50M tokens/month, the API is cheaper by $1,859.70/month.
† Anthropic's newest models (Fable 5, Opus 4.8, Sonnet 5) tokenize the same text into ~30% more tokens; the calculator corrects for this. Prices exclude batch (−50%) and prompt-caching discounts.
OpenAI pricing vs a rented H100 (July 2026)
| OpenAI model | Input $/1M | Output $/1M | Blended @4:1 | Break-even vs H100 rig |
|---|---|---|---|---|
| GPT-5.6 Sol | $5.00 | $30.00 | $10.00 | ≈211M tokens/mo |
| GPT-5.6 Terra | $2.50 | $15.00 | $5.00 | ≈422M tokens/mo |
| GPT-5.6 Luna | $1.00 | $6.00 | $2.00 | ≈1.05B tokens/mo |
| GPT-5.4 Mini | $0.75 | $4.50 | $1.50 | ≈1.41B tokens/mo |
| GPT-5.4 Nano | $0.20 | $1.25 | $0.41 | ≈5.1B — beyond one H100 |
The rig column assumes the preset above: a RunPod H100 80GB at $2.89/hr (≈$2,109.70/month) running an open-weights model at 1,500 tokens/sec and 60% utilization — an effective $0.90 per 1M tokens with a monthly capacity of about 2.33B tokens.
The math behind the table
blended $/1M = (4 × input price + output price) ÷ 5rig $/1M = $2,109.70 ÷ 2,332.8M tokens ≈ $0.90break-even = $2,109.70 ÷ blended $/1M
Worked example. A 500M tokens/month workload at 4:1 on GPT-5.6 Terra costs 500 × $5.00 = $2,500/month. The preset H100 rig costs $2,109.70/month and can produce ≈2.33B tokens, so self-hosting saves about $390/month — provided open-weights quality is acceptable and you run the server yourself. The same workload on GPT-5.4 Mini costs $750/month, and the API wins by roughly $1,360.
When the API wins — and when the H100 does
OpenAI wins below the break-even, and it is not close. Under ~200M tokens/month even the flagship tier beats a dedicated H100, because the server bills around the clock whether you use it or not. Spiky traffic makes this worse: an API bill scales down to zero on quiet days, a rental never does. Add frontier quality, function calling and zero operations work, and low-to-mid volume is simply API territory.
Self-hosting wins on sustained, high-volume, quality-tolerant work. From roughly 422M tokens/month against Terra — a volume real production apps do reach — a single H100 running Llama 3.3 70B undercuts the API, and it keeps getting better with scale until the capacity ceiling (~2.33B tokens/month at the preset assumptions). It is also the only option when data may not leave your infrastructure at all.
Count the costs the price list hides. Someone has to deploy vLLM, monitor it, handle model updates and be on call — the ops-overhead switch above exists because real-world estimates put total self-hosting cost at 2–5× the bare GPU rent. And 60% utilization is generous for a single-tenant box; drop it to 30% and every break-even volume doubles. Be honest with those two dials before migrating.
The pragmatic path is usually hybrid. Try OpenAI's Batch API first — its ~50% discount is the cheapest "self-hosting" you will ever run, with zero ops. If volume keeps growing, route the high-volume, low-stakes traffic to a self-hosted or hosted open-weights model and keep frontier calls on the API. Start by estimating your real monthly token usage, then compare all providers — not just OpenAI — in the main calculator.
Frequently asked questions
Related tools and guides
- LLM Token Cost Calculator — the main tool, with all providers and servers
- How to estimate your monthly LLM token usage — get the volume number this comparison depends on