Self-Hosted LLM vs API Break-Even: Tokens per Day by Model (Live GPU Prices)

"Should we run our own model?" usually gets answered with a GPU's hourly price and a frontier model's per-token price. That comparison flatters the GPU. The fair test is the same open model on someone else's GPU, billed per token, and those hosts are cheap.

So we priced both sides for four open models of different sizes. The GPU side uses aliteq's own tracker, which has logged Runpod's and Vast's prices every hour since 21 July 2026; this snapshot was captured at 19:21 UTC on 2 October 2026. For the API side we read each vendor's pricing page on 2 October 2026. The throughput figures are other people's published benchmarks. We didn't deploy or benchmark any of these models ourselves.

The break-even, model by model

One rented GPU beats the cheapest same-model API only if you push tens of millions of tokens through it every workday, and for three of our four models that's more than the card can serve. Against closed frontier models, break-even drops to between 2 and 11 million tokens a workday.

The formula is the GPU's monthly cost divided by what the API charges for the same tokens:

break-even tokens a workday = (GPU $/hour × 176 hours + $68 admin) ÷ (22 workdays × blended API price per token)

The blended price assumes eight input tokens for every output token, which is typical when you send documents and get short answers back: (8 × input price + output price) ÷ 9. For Llama 3.1 8B on DeepInfra that's (8 × $0.02 + $0.04) ÷ 9 = $0.022 per million tokens, so an RTX 4090 at $128 a month needs 128 ÷ (22 × 0.022) = 262 million tokens a workday. The ceiling column is what one card can process in an 8-hour day if it's perfectly busy, from the benchmarks below.

Three things jump out:

• The cheapest host sets the bar, and it's low. DeepInfra sells Llama 3.1 8B at $0.02 in and $0.04 out per million tokens. No rented card gets near that per token at a volume it can serve.
• gpt-6-luna is an open model's toughest rival. At $0.10 in and $0.50 out, it's cheaper than renting a GPU for gpt-oss-120b or Llama 3.3 70B at any volume one card can handle.
• Frontier models are where self-hosting wins. At $2 in and $10 out, gpt-6.1-sol and Claude Sonnet 5.5 cost enough that a $357-a-month GPU pays off from about 5.6 million tokens a workday. That's a quality decision, not just a hosting one: you're swapping a frontier model for an open one.

Which GPU each model needs, and what it rents for

Each model needs a card with enough memory for its weights plus the working memory for your context. The smallest card that fits a good-quality version runs from $0.34 an hour for Llama 3.1 8B to $1.64 an hour for gpt-oss-120b and Llama 3.3 70B, on Runpod on 2 October 2026.

The memory figures for gpt-oss-120b and Qwen3-Coder-30B come straight from our cost to run pages. The two Llama pages can't show working memory yet because Meta's repository is licence-gated, so we applied the same formula to the model shapes published in public copies of Meta's config files. Llama 3.3 70B also fits a 48 GB card such as an L40S at 4-bit, with little room to spare. Every card that fits each model, with live prices, is on the model pages: gpt-oss-120b, Qwen3-Coder-30B, Llama 3.1 8B and Llama 3.3 70B.

We used Runpod's price because it's one fixed price per card type. Vast.ai is a marketplace of individual hosts, so we show its median listing, not its cheapest one, which moves hour to hour. Live rates for every card are on our GPU price table, and the best GPU for each model pages rank cards by fit and price.

Referral link

Testing an open model before you commit?

An RTX 4090 that runs Llama 3.1 8B was listed at $0.34 an hour on Runpod on 2 Oct 2026. Rent by the hour, stop the pod when you're done. Prices move.

See live GPU pricesCompare live prices first →

Referral link — we may earn a commission at no cost to you. Prices on our compare page are the provider's live figures, cheapest first; this never changes the ranking.

How many tokens one GPU can actually serve

One GPU serving many users at once processes somewhere between about 1,000 and 4,600 tokens a second for these models, according to published benchmarks. That puts its ceiling at roughly 30 million to 130 million tokens in an 8-hour day, and you'll use only part of it.

These are ceilings measured with the card fully loaded. CloudRift's runs sent 400 requests at once, 1,000 tokens in and 1,000 out, and counted both directions; at that load each user waits a long time for the first word. Real traffic comes in peaks, so we treat 50% to 70% of the ceiling as usable. The two benchmark vendors rent GPUs, and Spheron's figure is a rounded one; both are published by companies that sell the self-hosting side.

One more caveat on the big card. The benchmarks used full-power RTX PRO 6000s, and the cheapest one on Runpod is the Max-Q, a 300-watt version that may be slower. Runpod's full-power card was $1.69 an hour, which moves the break-even figures up by about 2.5%.

Same model, different host: the price gap is huge

The same open model can cost ten times more on one API than on another, so who you compare against decides whether self-hosting looks smart. For Llama 3.3 70B, self-hosting beats Together's price at 15.6 million tokens a workday but never beats DeepInfra's.

Two details matter when you read that table:

• The cheap Llama endpoints are quantized. DeepInfra's Turbo versions of Llama run at 8-bit (FP8), and the self-hosted figures in this article assume 4-bit to 8-bit too. Neither is the full-precision model, and quality differences are small but not zero.
• Claude's newer models count more tokens. Anthropic says Claude 4.7 and later models produce about 30% more tokens for the same text, so Sonnet 5.5 costs more per word than its per-token price suggests. Haiku 4.5 uses the older tokenizer. Measured in words rather than tokens, self-hosting breaks even sooner against Sonnet 5.5 than the table shows.

Groq also lists Llama 3.1 8B and Llama 3.3 70B as enterprise models with no public price, and Fireworks prices models it doesn't list individually by size. We only used prices printed for the exact model.

What it costs at three volumes

At 1 million or 10 million tokens a workday, every API in this article costs less than renting a GPU for gpt-oss-120b. At 50 million, the rented card undercuts Claude and gpt-6.1-sol, but still not gpt-6-luna or any API for gpt-oss-120b itself.

How big is 10 million tokens a day? A 25-person team asking an internal assistant 20 questions a day each, with about 4,000 tokens of documents and question going in and 500 coming back, uses about 2.25 million. So 1 million a workday is light team use, and 50 million is a product with real customers.

For small models, the fixed cost isn't the GPU. Our admin line is one hour a month at $68, the loaded hourly cost of a US systems administrator from the Bureau of Labor Statistics median wage. That's more than the $60 of RTX 4090 time a month.

The assumptions, and what they leave out

Every number above rests on choices we made. Change them and the break-even moves, but the pattern holds: the cheapest same-model API is hard to beat, and frontier prices are easy to beat.

• Hours. Eight hours a day, 22 workdays, with the pod stopped at night: 176 GPU-hours a month. Run it around the clock and the GPU costs about 3.5 times as much; break-even against Sonnet 5.5 rises to 14.4 million tokens a day for gpt-oss-120b, and against the cheapest same-model API it's still out of reach for all four models except Qwen3-Coder-30B, which would need 204 million a day, about half its ceiling.
• Token mix. Eight input tokens per output token. Chat with long answers has more output, which makes APIs dearer and helps the GPU a little.
• Admin. One hour a month. A first deployment of vLLM or another serving stack takes far longer than that; we didn't price setup time.
• Left out on purpose. Storage, data transfer, idle time while a model loads, a second GPU for failover, and prompt caching and batch discounts on the API side. All but the last make self-hosting more expensive.
• Throughput. Published benchmarks from GPU sellers, not our tests, at full load. Your traffic pattern decides how much of that you get.

If you're thinking about buying the hardware instead of renting it, the three-year math is different. Our calculator compares an owned AI server with renting the same card:

The full write-up is in AI server vs cloud GPU: the 3-year cost.

When self-hosting does make sense

Self-hosting pays when it replaces a frontier model at a few million tokens a day, or when the reason isn't price. Here's how to decide.

If keeping data inside the EU is the reason, the routes and prices are different; we cover them in private LLM cost for EU companies. More cost guides for running AI in a business are on the AI Automation hub.