Self-hosting an LLM costs less than a premium API once the volume is steady and high enough. A dedicated server with a 96 GB GPU at $1,656.15 a month breaks even against a model priced at $2 per million input tokens and $10 per million output tokens at about 414 million tokens a month, and against a flagship model at $10 and $50 at about 83 million. Below the break-even volume, paying per token costs less.
So the answer to "is it cheaper to self-host an LLM?" depends on your monthly volume and on which model you would otherwise pay for. MassiveGRID sells GPU dedicated servers, so we have an interest here, and this post still shows the cases where an API is the better fit. All API prices were checked on 16 November 2026.
What do LLM APIs cost per million tokens in November 2026?
The table lists standard pay-as-you-go list prices for the premium models of the three large vendors, in US dollars per one million tokens, as shown on each vendor's pricing page on 16 November 2026. These are the models a self-hosted deployment is usually weighed against. The last column is a blended price for a workload with three input tokens for every output token, which is typical for chat with history or retrieval: (3 x input + output) / 4.
| Vendor | Model | Input | Output | Blended at 3:1 |
|---|---|---|---|---|
| OpenAI | gpt-6-astra | $10.00 | $50.00 | $20.00 |
| Anthropic | Claude Fable 5.2 | $8.00 | $40.00 | $16.00 |
| Anthropic | Claude Opus 5 | $5.00 | $25.00 | $10.00 |
| Anthropic | Claude Opus 5.5 | $4.00 | $20.00 | $8.00 |
| gemini-3.1-pro-preview (prompts to 200k tokens) | $2.00 | $12.00 | $4.50 | |
| OpenAI | gpt-6.1-sol | $2.00 | $10.00 | $4.00 |
| Anthropic | Claude Sonnet 5.5 | $2.00 | $10.00 | $4.00 |
Notes on the figures. OpenAI's prices are the standard tier for prompts up to 272K tokens, read from its developer pricing documentation. Cached input is billed at a lower rate ($0.10 per million on gpt-6.1-sol), which lowers the blended price for prompts with a long repeated prefix, so use your real invoice where you have one. gemini-3.1-pro-preview costs $4.00 and $18.00 above 200k-token prompts, a blended $7.50. Anthropic lists Claude Sonnet 5 at the same $2 and $10 as Sonnet 5.5.
How do you calculate the break-even point?
A dedicated server costs the same whether it serves one request or a million. An API costs nothing when idle and scales with every token. The break-even is the monthly volume at which the two bills are equal:
blended price per 1M tokens = (r x input price + output price) / (r + 1)
break-even tokens per month = server price per month / blended price x 1,000,000
Here r is your ratio of input tokens to output tokens. Measure it from your own logs: an agent with long tool outputs can be 10:1 or more, a writing assistant can be close to 1:1. Because output tokens cost four to five times as much as input tokens on most APIs, the ratio moves the result a lot.
For the two single-GPU MassiveGRID servers, at a 3:1 ratio:
| API model replaced | Blended price per 1M | Break-even at $449.99 (24 GB server) | Break-even at $1,656.15 (96 GB server) |
|---|---|---|---|
| gpt-6-astra | $20.00 | 22.5 million tokens | 82.8 million |
| Claude Fable 5.2 | $16.00 | 28.1 million | 103.5 million |
| Claude Opus 5 | $10.00 | 45.0 million | 165.6 million |
| Claude Opus 5.5 | $8.00 | 56.2 million | 207.0 million |
| gemini-3.1-pro-preview | $4.50 | 100.0 million | 368.0 million |
| gpt-6.1-sol or Claude Sonnet 5.5 | $4.00 | 112.5 million | 414.0 million |
Below the break-even volume in a row, paying per token costs less. Above it, the fixed monthly price is the lower bill.
Two corrections make the comparison fair. First, the servers have a one-time setup fee ($346.48 and $1,656.15), so in the first year the effective monthly cost is $478.86 and $1,794.16 and every break-even rises by 6% to 8%. Second, the rows compare different things: an open model on your own card is not gpt-6-astra or Claude Fable 5.2. The break-even is only meaningful if an open model does the job well enough, which is a question to answer with your own evaluation set before any cost calculation.
Three worked examples
1. A support assistant: 240 million tokens a month. 2,000 conversations a day, each with 3,000 input tokens (instructions, retrieved articles, history) and 1,000 output tokens, is 8 million tokens a day. On gpt-6.1-sol that is 180 million input tokens at $2.00 plus 60 million output tokens at $10.00: $960 a month. On gemini-3.1-pro-preview it is $1,080 and on Claude Opus 5.5 $1,920. The 96 GB server at $1,656.15 costs $696.15 more than gpt-6.1-sol and $263.85 less than Claude Opus 5.5. The 24 GB server at $449.99 costs less than all three, if a model that fits in 24 GB, such as a 27B or 32B model at 4-bit, answers your tickets well enough.
2. A document pipeline: 1 billion tokens a month. 750 million input and 250 million output tokens cost $4,000 on gpt-6.1-sol, $4,500 on gemini-3.1-pro-preview and $8,000 on Claude Opus 5.5. A 96 GB server running Llama 3.3 70B (a 43 GB download at the default 4-bit quantization on Ollama) costs $2,343.85 a month less than gpt-6.1-sol and $6,343.85 less than Claude Opus 5.5, or $2,205.84 and $6,205.84 in the first year with the setup fee spread over twelve months.
3. A coding agent on a frontier model: 100 million tokens a month. 75 million input and 25 million output tokens on Claude Fable 5.2 cost $600 plus $1,000: $1,600 a month, about the price of the 96 GB server. On paper that is break-even. In practice the server would run a different, smaller model, so the real question is whether the open model completes the same tasks. If it needs twice as many attempts, the saving is gone.
How many tokens can one GPU server actually serve?
A break-even volume only matters if the card can produce it. We publish no benchmarks of our own, so the bounds below come from third-party tests, each with its setup. None of them used the 300 W Max-Q edition of the RTX PRO 6000 that our servers carry; NVIDIA rates the Max-Q at 3,511 AI TOPS against 4,000 for the 600 W Workstation Edition, so expect somewhat less.
- CloudRift, 9 October 2025: one RTX PRO 6000 serving Llama 3.3 70B (AWQ INT4) with vLLM reached 1,031 tokens per second with 400 concurrent requests of 1,000 input and 1,000 output tokens. The same card reached 8,425 tokens per second on Qwen3-Coder-30B-A3B (AWQ 4-bit).
- Akamai, 30 October 2025 (a cloud vendor that sells this GPU): one RTX PRO 6000 Blackwell Server Edition serving a 49B Nemotron model in FP8 through NVIDIA NIM, with 200 tokens in and 200 out, produced 27.42 tokens per second for a single request and 2,256.3 tokens per second in total at 100 concurrent requests.
- Single stream with Ollama: Database Mart reports about 134 tokens per second for gpt-oss 120b and about 185 for gpt-oss 20b on an RTX PRO 6000 (figure taken from a search snippet of its benchmark page).
A month has 2,628,000 seconds. If the CloudRift figure is generated tokens and the card were saturated every second of the month, 1,031 tokens per second is 2.71 billion output tokens, with the same number of input tokens in that test. At gpt-6.1-sol list prices that traffic would cost $5,419 plus $27,095: $32,514. The $1,656.15 server therefore matches that API's price at about 5% of the saturated throughput in that test. Real endpoints run at a fraction of saturation, which is why the break-even volume matters more than the ceiling.
For the three examples above capacity is not the constraint: 250 million output tokens a month is an average of 95 tokens per second.
To measure your own case, run your real prompts through a vLLM endpoint with realistic concurrency and record tokens per second and time to first token. What determines the result is the model size and quantization, the context length, the number of concurrent requests the engine can batch, and the card's memory bandwidth.
What about per-seat AI products and speech-to-text APIs?
Not every AI bill is metered in tokens. Assistants are often bought per user, and transcription per hour of audio. The same division gives a break-even for those: the server price divided by the price per seat or per audio hour. All prices below are list prices on 16 November 2026.
| Per-seat product | List price per user per month | Seats that equal $449.99 | Seats that equal $1,656.15 |
|---|---|---|---|
| ChatGPT Business, Standard seat (reported from OpenAI's pricing and help pages) | $25 monthly, $20 billed annually | 18 | 67 |
| Claude Team, Standard seat | $25 monthly, $20 billed annually | 18 | 67 |
| Microsoft 365 Copilot, enterprise add-on | $30.00 billed annually, $31.50 monthly | 15 | 56 |
| Gemini Enterprise, Business edition (reported from Google Cloud's page) | from $21 | 22 | 79 |
| GitHub Copilot Business | $19 | 24 | 88 |
| GitHub Copilot Enterprise | $39 | 12 | 43 |
The seat counts use the first price in each row and are rounded up. Read them with care: a seat buys a frontier model, the vendor's apps and its integrations, while a server runs an open model behind an interface you install yourself, such as Ollama with Open WebUI. Whether one card serves 67 or 88 people depends on how many of them type at the same moment, which you measure rather than assume. Microsoft 365 Copilot also needs a qualifying Microsoft 365 licence underneath.
| Speech-to-text API | List price | Per audio hour | Audio hours a month that equal $449.99 |
|---|---|---|---|
| Google Speech-to-Text V2, standard recognition | $0.016 per minute | $0.96 | 469 |
| AWS Transcribe, streaming (rate used in AWS's examples) | $0.01 per minute | $0.60 | 750 |
| OpenAI Whisper or gpt-4o-transcribe | $0.006 per minute | $0.36 | 1,250 |
| AWS Transcribe, batch (rate used in AWS's examples) | $0.006 per minute | $0.36 | 1,250 |
| Deepgram Nova-3, pre-recorded, pay as you go | $0.0043 per minute | $0.258 | 1,744 |
| AssemblyAI Universal-3.5 Pro, pre-recorded | $0.21 per hour | $0.21 | 2,143 |
| ElevenLabs Scribe v3 | $0.20 per hour | $0.20 | 2,250 |
A month has 730 hours, so 1,250 audio hours a month means transcribing about 1.7 hours of audio per hour of wall-clock time, around the clock. Whether a 24 GB card running an open Whisper model reaches that depends on the model size, the batch size and the audio itself; the self-hosted Whisper guide shows how to measure it. Below the break-even hours, paying per audio hour costs less.
Hourly GPU rental sits between a metered API and a monthly server. If you need your own endpoint for a few hours a day, the dedicated versus hourly GPU pricing comparison has the break-even hours.
Where is an API the better fit?
- Very low or uneven volume. Below the break-even volume in the table above, paying per token costs less, and idle months cost nothing. A pilot, an internal tool with a handful of users or a seasonal workload belongs on an API.
- Frontier quality. The largest proprietary models are not available as weights. The newest open flagships are released as weights but are far too large for one server: Mistral Large 4 is 1,060.6 GB and DeepSeek V4-Pro 864.7 GB on Hugging Face. If the task needs that class of model, the API is the practical route.
- No operations. An API has no drivers, no model updates, no monitoring and no on-call. A dedicated server is a single machine with a clean operating system: you install the stack, you keep it patched, and redundancy and backups are your job.
- Burst capacity. An API absorbs a tenfold spike. One card queues.
Where does self-hosting win?
Above the break-even volume the server is the lower bill. Below it, the case for your own server rests on these four things.
- Privacy. Prompts, documents and outputs stay on a machine that only you have root on. Nothing is sent to a model vendor, so no third-party AI provider processes the content, which matters most for client-confidential and regulated work. Private AI hosting covers this in detail.
- Control. You choose the model and the version, and it does not change or retire on someone else's schedule. Google's price list shows why that matters: gemini-3.8-flash doubles in price on 1 January 2027, and two Gemini 2.5 models are deprecated and rerouted from 1 December 2026.
- Fixed cost. The bill is $449.99 or $1,656.15 whether you send 10 million tokens or 2 billion, with unlimited traffic on the 1 Gbps port. Batch jobs, evaluations and experiments have no marginal cost, which changes how freely a team uses the model.
- Custom models. Fine-tuned weights, uncommon open models and your own adapters run on your card without a per-model hosting agreement. Unsloth's published minimum for 4-bit QLoRA fine-tuning of a 70B model is 41 GB, which fits a 96 GB card.
If two or more of these matter, the cost question becomes "is the premium acceptable?" rather than "is it cheaper?". At 240 million tokens a month, the premium of the 96 GB server over gpt-6.1-sol is $696 a month.
A practical way to decide
- Export a month of usage from your API dashboard: input tokens, output tokens, and peak requests per minute.
- Compute your blended price with the formula above and divide the server price by it.
- Test an open model on your own tasks before you commit to hardware. Run your evaluation set against a 32B and a 70B model for a few hours on a GPU rented by the hour, and compare the answers with what the API model gives you.
- Pick the memory class. Models up to about 32B at 4-bit fit 24 GB; 70B-class and 120B mixture-of-experts models need 96 GB. The guide to GPUs for LLM inference has the full table.
- Keep the API as a fallback. vLLM exposes OpenAI-compatible endpoints, so routing overflow or the hardest requests to an API is a configuration change. Workflow tools can be pointed at either: see running n8n and Flowise on a local LLM.
If the result favours your own hardware, the LLM hosting page maps models to the four servers.
Frequently Asked Questions
Is it cheaper to self-host an LLM than to use the OpenAI API?
At steady, high volume, yes. Against gpt-6.1-sol ($2.00 input, $10.00 output per million tokens) a $1,656.15 server breaks even at about 414 million tokens a month at a 3:1 input-to-output ratio, and a $449.99 server at about 112.5 million. Against gpt-6-astra at $10.00 and $50.00 the break-even is 82.8 million and 22.5 million tokens a month. Below the break-even volume, paying per token costs less.
How many tokens per month justify a dedicated GPU server?
Divide the server's monthly price by your blended API price per million tokens. As a rule of thumb for the 96 GB server from the 16 November 2026 list prices: about 80 to 100 million tokens a month against flagship models at $8 to $10 input, about 200 million against a $4/$20 model and about 400 million against $2/$10 models. Add 6% to 8% in the first year for the setup fee on our GPU servers.
How many ChatGPT Business or Microsoft 365 Copilot seats cost the same as a GPU server?
At the list prices on 16 November 2026, $1,656.15 a month equals 67 seats at $25 (ChatGPT Business as reported from OpenAI's pages, or Claude Team, both billed monthly) or 56 Microsoft 365 Copilot seats at $30.00, and $449.99 equals 18 or 15 seats. A seat includes a frontier model and the vendor's apps, while the server runs an open model behind an interface you install, so compare what users actually need before comparing the bills.
What does self-hosting cost besides the server?
The one-time setup fee ($346.48 or $1,656.15 on the two single-GPU servers), the time to install and maintain drivers and the serving stack, and your own backups and redundancy, since a dedicated server is a single machine. Traffic is included: the servers have a 1 Gbps port with unlimited traffic.
LLM hosting on your own GPU server
When privacy, a fixed bill or custom models decide the question, MassiveGRID GPU dedicated servers give you the whole machine: 24 GB of GPU memory for $449.99 per month for models up to about 32B, 96 GB from $1,656.15 for 70B-class models, or 384 GB across four cards for $5,966.44. Each has a 1 Gbps port with unlimited traffic, root access and a clean Linux install, so prompts and outputs stay on hardware only you control.
See LLM HostingFurther Reading
- OpenAI API pricing (standard tier, per 1M tokens, and transcription rates)
- Anthropic Claude API pricing
- Google Gemini API pricing (paid tier, including the 1 January 2027 change)
- Claude plans pricing (Team seats)
- ChatGPT Business pricing
- Microsoft 365 Copilot enterprise pricing
- Gemini Enterprise (Google Cloud)
- GitHub Copilot plans
- Google Cloud Speech-to-Text pricing
- Amazon Transcribe pricing
- Deepgram pricing (Nova-3)
- AssemblyAI pricing
- ElevenLabs API pricing (Scribe speech-to-text)
- CloudRift: benchmarking RTX GPUs for LLM inference with vLLM (9 October 2025)
- Akamai: benchmarking the RTX PRO 6000 Blackwell Server Edition (30 October 2025)