Frontier models or your own GPUs? What each costs in 2026
Closed frontier models still lead on capability but open-weight models you can run yourself are closing the gap. Here is what each option might cost a large company in November 2026: API tokens, rented GPUs and owned hardware, priced side by side.
by David Bunting, Founder
Three ways to buy the same capability
Every large company using AI at scale ends up asking the same question: keep paying a frontier lab by the token, or run a model ourselves? In practice there are three options:
- Rent the model. Pay Anthropic, OpenAI or Google for their closed frontier models.
- Rent the hardware. Run an open-weight model on GPUs from a cloud provider.
- Own the hardware. Run an open-weight model in your own data centre.
Each option has pros and cons. This piece sets out what each one costs today and what you get for the money.
About the numbers
How close are the open models?
The open-weight models worth considering for frontier-class work are large mixture-of-experts models, mostly from Chinese labs: Moonshot’s Kimi, Z.ai’s GLM and DeepSeek. Whilst they are highly capable, across the board, they are not yet as good as the best closed models.
On aggregate indices the gap is still wid. BenchLM’s overall score puts Claude Fable 5.2 22.87 points ahead of the best open-weight model on a 100-point scale. The Respan Index has Claude Opus 5.5 at 74.0 against 59.4 for the best open model, DeepSeek V4.1 Flash.
On specific tasks the picture is more mixed. On Vals AI’s Terminal-Bench 4.0, an agentic coding benchmark, Claude Fable 5.2 and Opus 5.5 share the lead at 65.15%. Moonshot’s Kimi K3.1 Pro, whose weights are on Hugging Face, is close behind at 63.64%. GLM-5.3 scores 38.89%. Which number matters depends on what you need the model to do, so the only benchmark that settles it is your own workload.
Size matters as much as quality, because size sets the hardware bill:
| Model | Size | Weights | Licence |
|---|---|---|---|
| Kimi K3 (Moonshot) | 2.8T parameters, 104B active | 1.56 TB | Kimi K3 License |
| GLM-5.3 (Z.ai) | 753B parameters | 756 GB | GLM-5.3 License |
| DeepSeek V4.1 Flash | 552B backbone plus 196B memory, 8–16B active | 510 GB | MIT |
The licences differ more than the word “open” might suggest. DeepSeek’s is plain MIT, whilr Kimi K3’s exempts internal use, but a business selling access to the model needs a separate agreement once revenue passes $20 million over 12 months, and “very large“ commercial products must display the Kimi name. GLM-5.3’s asks the same kind of business for a security review by Z.ai, but only above $10 billion in revenue. For internal use at most companies, these are legitimate legal concerns.
Option one: frontier models on an enterprise contract
This is where most large companies start. You buy seats for staff, API access for developers, or both. Current API prices:
| Model | Input, per 1M tokens | Output, per 1M tokens |
|---|---|---|
| GPT-6 Astra | €9.10 | €45.50 |
| Claude Fable 5.2 | €7.28 | €36.40 |
| Claude Opus 5.5 | €3.64 | €18.20 |
| Gemini 4 Carbon (from 1 December) | €3.64 | €18.20 |
| Claude Sonnet 5.5 | €1.82 | €9.10 |
Seats are simpler to budget but harder to compare. Claude Enterprise lists at about €18 a seat a month, billed annually with a 20-seat minimum, and usage on top at API rates. ChatGPT Enterprise is priced per user per month on annual plans, but OpenAI doesn’t publish the figure, you can only get it from sales.
In its favour:
- The best models available. On every index above the top closed models still lead.
- No initial large capital spend and nothing to operate.
- New models arrive without a migration project. Upgrading usually means changing a model name.
- Contractual terms on data use and retention, plus the admin controls a large organisation expects.
Against it:
- Spend grows with use. Only your own budgets cap it.
- Prices move on the vendor’s timetable. Gemini 4 Carbon’s introductory price ends on 30 November, and from 1 December it doubles. It moves the other way too: Anthropic launched Sonnet 5 at an introductory price, then made that the standard price.
- Control of the data is contractual, not physical. Prompts and documents travel to the vendor’s infrastructure, and for some data classes that ends the discussion.
- You depend on one vendor’s roadmap, model retirements and terms.
Option two: open models on rented GPUs
The middle path. You pick an open-weight model, rent GPUs from a cloud provider, and serve the model with an inference engine such as vLLM or SGLang.
The model sets the hardware. A node of eight NVIDIA H200s has 1,128 GB of GPU memory. GLM-5.3 and DeepSeek V4.1 Flash fit on one, with room left for the working memory each conversation needs. Kimi K3 does not: its weights alone are larger than the node, so you need two nodes or newer, larger GPUs, at roughly double the cost.
Rental prices vary widely. IONOS lists eight H200 PCIe GPUs at about €23.74 an hour. RunPod’s Secure Cloud lists one H200 at about €4.81 an hour, so eight cost €38.48.
In its favour:
- You choose the model, and can switch when a better one ships without renegotiating a contract.
- You choose the provider and region, so you can match your residency requirements.
- No capital spend. Billing is by the hour, not by the token, so heavy use costs the same as light use.
Against it:
- You pay for idle hours. A node left running overnight costs the same as a busy one, and scaling it down takes engineering.
- You run it: the inference engine, model upgrades, monitoring, security patches and capacity for peak load.
- Quality is capped at what the open models can do.
- The data still sits in someone else’s data centre, under a contract.
Option three: open models on your own hardware
The same software, on servers you buy. Mercatus put an eight-GPU HGX H200 server at €291,000 to €382,000 in July. AMCompute reported about €309,000 for a volume order in October. Spread over three years, the Mercatus range comes to €8,100 to €10,600 a month.
Then there is power. NVIDIA rates the DGX H200 at 10.2 kW maximum. Flat out for a month, that is about 7,450 kWh. At Germany’s average industrial price of 12.7 cents per kWh in February 2026, it comes to about €950 a month before cooling. Your tariff will differ.
In its favour:
- Physical control. The data never leaves your building, and the system can run with no internet connection at all.
- The lowest cost per hour, as long as the hardware is busy.
- Predictable cost. Once the hardware is bought, heavy use costs little more than light use.
Against it:
- Capital spend up front, and lead times on top.
- Obsolescence. NVIDIA now ships a new data centre GPU generation roughly every year, and the H200 is already behind the B200 and B300. Three-year depreciation assumes the hardware is still worth running in year three.
- Staff. Someone has to run the servers, the models and the security around them, and those people are hard to hire.
- You own capacity planning: too little and users queue, too much and idle hardware is costing you money.
The same money, three ways
To compare like with like, take one eight-H200 node, enough for GLM-5.3 or DeepSeek V4.1 Flash, and price a month of round-the-clock use each way:
| Eight H200 GPUs, running all month | Per month | Basis |
|---|---|---|
| Rented from IONOS | ≈ €17,300 | €23.74 an hour for eight H200 PCIe GPUs, 730 hours |
| Rented from RunPod Secure Cloud | ≈ €28,100 | €4.81 per GPU an hour, eight GPUs, 730 hours |
| Owned, over three years | ≈ €9,000–11,600 | €291,000–382,000 over 36 months, plus 7,450 kWh at 12.7 cents. Excludes cooling, space and staff. |
Then ask what the cheapest rental figure, €17,300 a month, buys from the frontier APIs instead:
| €17,300 a month spent on | Tokens per month |
|---|---|
| Claude Sonnet 5.5 | ≈ 6.3 billion |
| Claude Opus 5.5 or Gemini 4 Carbon | ≈ 3.1 billion |
| Claude Fable 5.2 | ≈ 1.6 billion |
| GPT-6 Astra | ≈ 1.3 billion |
Four variables matter in the end:
- Utilisation. The rental figures assume the node runs all month. If you only need it twelve hours a day and switch it off for the rest, IONOS comes to about €8,700, less than owning. Against that price, owning only comes out ahead if the node is busy for more than about 12 to 16 hours a day.
- Volume. €17,300 buys about 3.1 billion tokens a month on Opus 5.5. That is roughly 140,000 tokens a working day for each of 1,000 employees. Well below that, the API costs less than any node. Well above it, self-hosting starts to pay, provided one node can keep up.
- Capacity.Whether one node can serve your peak load depends on how many people use it at once and how long their prompts are. I haven’t found independent throughput benchmarks for these models that I would rely on, so test with your own workload before sizing anything.
- People. None of the self-hosted figures include the engineers who run them. For one node that may be part of a rolebut for a fleet that could be a whole team.
What decides it
There is no single right answer. The questions that decide it for a given company:
- How many tokens do you use a month, and how steady is that? Spiky use suits paying per token. Steady, heavy use suits hardware.
- What quality does each workload need? Test the open models on your own tasks before assuming they are good enough, or that they aren’t.
- Which data can leave your network, and under what contract? Some data classes settle the question on their own.
- Who will run it? Self-hosting is an operations commitment, not a once-off purchase.
- How comfortable are you with hardware cycles? Buying GPUs is a bet on what they will be worth in three years.
Prices in this market move every month. Treat the numbers above as a snapshot from mid-November 2026, and redo the arithmetic with your own quotes.
Sources and assumptions
Prices are list prices on 16 November 2026. Dollar prices are converted at 1 USD = €0.91. A month is 730 hours. Token volumes assume 150,000 output tokens for every million input tokens. Hardware is depreciated evenly over 36 months, and power is costed at the DGX H200’s 10.2 kW maximum.
- Anthropic API pricing
- Claude plans
- OpenAI API pricing
- ChatGPT plans
- Gemini API pricing
- BenchLM open-weight statistics
- Respan Index
- Vals AI Terminal-Bench 4.0
- Kimi K3 model card
- Kimi K3.1 Pro on Hugging Face
- GLM-5.3 model card
- DeepSeek V4.1 Flash model card
- IONOS GPU VM pricing
- RunPod pricing
- Mercatus, HGX H200 server price (July 2026)
- AMCompute, H200 server values (October 2026)
- NVIDIA DGX H200 specifications (PNY)
- Intratec, German industrial electricity price