Short answer
At low volume an LLM API is far cheaper than self-hosting: 10 million input and 2 million output tokens a month cost $2 on a small model and $40 on a mid-tier model at list price, while one cloud GPU running all month costs about $588 to $5,095 before staff. At Apex36, a private LLM deployment starts at $8,000.
Small models, per 1M tokens
$0.10 / $0.50
Claude Haiku 5.5, GPT-6 Luna
Mid-tier, per 1M tokens
$2 / $10
Claude Sonnet 5.5, GPT-6.2 Sol
One GPU, 24/7, a month
$588 to $5,095
L4 to H100, on demand
Apex36 private LLM
From $8,000
per deployment, in your cloud
How much does the OpenAI API cost in 2026?
As of 18 Nov 2026, the OpenAI API costs $0.10 per million input tokens on GPT-6 Luna, $2 on GPT-6.2 Sol and $8 on GPT-6.2 Astra, with output tokens at five times the input price (openai.com/api/pricing). Standard list prices per million tokens:
| Model | Input per 1M tokens | Output per 1M tokens | Source |
|---|---|---|---|
| GPT-6 Luna | $0.10 | $0.50 | openai.com/api/pricing |
| GPT-6.2 Sol | $2 | $10 | openai.com/api/pricing |
| GPT-6.2 Astra | $8 | $40 | openai.com/api/pricing |
Worked example: 10 million input and 2 million output tokens a month cost $2 on GPT-6 Luna, $40 on GPT-6.2 Sol and $160 on GPT-6.2 Astra.
How much does the Claude API cost in 2026?
As of 18 Nov 2026, the Claude API costs $0.10 per million input tokens on Claude Haiku 5.5, $2 on Claude Sonnet 5.5 and $4 on Claude Opus 5.5, with output tokens at five times the input price (Anthropic's pricing page). Standard list prices per million tokens:
| Model | Input per 1M tokens | Output per 1M tokens | Source |
|---|---|---|---|
| Claude Haiku 5.5 | $0.10 | $0.50 | platform.claude.com pricing (prompts up to 100k tokens) |
| Claude Sonnet 5.5 | $2 | $10 | platform.claude.com pricing |
| Claude Opus 5.5 | $4 | $20 | platform.claude.com pricing |
Worked example: the same 10 million input and 2 million output tokens cost $2 on Claude Haiku 5.5, $40 on Claude Sonnet 5.5 and $80 on Claude Opus 5.5.
How much does the Gemini API cost in 2026?
As of 18 Nov 2026, the Gemini API costs $0.30 per million input tokens on Gemini 3.5 Flash-Lite and $2 on Gemini 4 Carbon, which rises to $4 on 1 Dec 2026 (ai.google.dev pricing). Output costs $2.50 and $10 per million tokens, with Carbon's output also doubling to $20.
| Model | Input per 1M tokens | Output per 1M tokens | Source |
|---|---|---|---|
| Gemini 3.5 Flash-Lite | $0.30 | $2.50 | ai.google.dev pricing |
| Gemini 4 Carbon | $2 until 30 Nov 2026, then $4 | $10, then $20 | ai.google.dev pricing |
All prices are standard rates per million tokens. Batch processing is 50% off at Anthropic and OpenAI; long prompts and data residency change prices. Check the provider's page before relying on a number.
How much does a cloud GPU cost to run an LLM?
One on-demand cloud GPU costs about $588 a month (AWS L4) to $5,095 a month (Azure H100 NVL) running 24/7, before staff and redundancy (AWS and Azure price feeds, November 2026). A rented GPU bills for every hour it runs, busy or idle. Prices per instance, Linux, us-east-1 and eastus, bare compute only:
| Instance | Relative cost | A month, 24/7 |
|---|---|---|
| AWS g6.xlarge1× L4 · $0.8048/hr | $588 | |
| AWS g5.xlarge1× A10G · $1.006/hr | $734 | |
| AWS g6e.xlarge1× L40S · $1.861/hr | $1,359 | |
| Azure NC24ads A100 v41× A100 80 GB · $3.673/hr | $2,681 | |
| AWS p5.4xlarge1× H100 · $6.88/hr | $5,022 | |
| Azure NC40ads H100 v51× H100 NVL · $6.98/hr | $5,095 |
Sources: AWS EC2 on-demand pricing (us-east-1) and Azure retail prices (eastus).
When does a self-hosted LLM cost less than an API?
A self-hosted LLM costs less than an API only at high, steady volume: at mid-tier API prices, one L40S GPU at $1,359 a month breaks even near 400 million tokens a month, before overhead. Self-hosted LLM cost is the GPU's hourly rate × 730 hours, multiplied by 1.7 to 2.7 for redundancy, monitoring and people (DigitalOcean). Two formulas cover the decision:
API cost
(input tokens × input price) + (output tokens × output price). You pay only for what you send.
Self-host cost
Hourly rate × 730 hours, then × 1.7 to 2.7 for redundancy, monitoring and people (DigitalOcean).
~400M
tokens a month
Illustrative example: at mid-tier API prices (about $3.33 per million tokens blended), one L40S instance at $1,359 a month breaks even near 400 million tokens a month, before any overhead. DigitalOcean's rule is to self-host "only if break-even falls under ~30% utilization OR there is no per-token offering available for your model". If data must stay in your environment, see our private LLM service.
When is self-hosting an LLM worth it?
Self-hosting an LLM is worth it in four cases: your data must stay in your own environment, you need a fine-tuned or custom model, your volume keeps a GPU busy all month, or the model has no API offering.
Data can't leave
Your data must not leave your own environment.
Custom model
You need a fine-tuned or custom model.
High, steady volume
Volume is high and steady enough to keep a GPU busy.
No API offering
The model you want has no API offering.
Outside those cases, a cheap API model often beats a rented GPU outright.
Can you keep LLM data private without running GPUs?
Yes: managed models served inside your own cloud account keep LLM traffic private without running GPUs. Claude runs on AWS Bedrock, OpenAI models on Azure and Gemini on Vertex AI; you pay per token, and the data stays in the cloud account you control.
What Apex36 charges
From $8,000 per deployment. One use case, one model set up in your cloud account and an evaluation on your data. Running costs are billed to your own cloud and model accounts.
From
$8,000
per deployment
Frequently asked questions
How much does the OpenAI API cost?
It's priced per million tokens, and input and output are priced separately. As of the date checked, GPT-6 Luna is $0.10 in and $0.50 out, GPT-6.2 Sol is $2 and $10, and GPT-6.2 Astra is $8 and $40. Batch processing is half price. 10 million input and 2 million output tokens a month cost $2 on Luna or $40 on Sol.
Is it cheaper to self-host an LLM or use an API?
At low or uneven volume, the API is much cheaper. A rented GPU bills for every hour, and the real cost is 1.7 to 2.7 times the bare GPU price once you add redundancy, monitoring and people. Self-hosting pays with high, steady use.
How much does it cost to run an LLM on your own GPU?
One on-demand cloud GPU costs about $588 a month (L4) to $5,095 a month (H100 NVL) running all month on AWS or Azure, before staff and redundancy.
When should a company self-host?
When data must stay in your environment, when you need a fine-tuned or custom model, or when volume is high and steady enough to keep a GPU busy.
Can we keep data private without running GPUs?
Often yes: managed models served inside your own cloud account (Claude on AWS Bedrock, OpenAI models on Azure, Gemini on Vertex AI) keep traffic in your cloud with per-token pricing.
What does Apex36 charge to set up a private LLM?
From $8,000 per deployment: one use case, one model setup in your cloud account and an evaluation on your data. Model and cloud costs are billed to your accounts.
Sources
The market figures in this guide come from 6 public sources, read on and rechecked monthly. Each figure names its source and says whether it is an aggregate across many firms or one firm's estimate. Apex36 sells this service; our own prices are on /pricing.