Self-Hosted LLM vs API: Where the Cost Crossover Actually Sits in 2026

Every team burning through an AI budget eventually asks the same question: does a self-hosted LLM vs API comparison end with cheaper tokens on your own hardware, or does the subscription stay the rational choice? The honest answer is that the crossover exists but it is workload-shaped, not universal. Cloud APIs bill per token - GPT-5 runs $1.25 per million input and $10 per million output tokens, Claude Sonnet 5 runs $2/$10, and Gemini 2.5 Flash runs $0.30/$2.50 - while self-hosting trades a one-time hardware bill plus electricity for tokens that cost almost nothing at the margin. This guide works the arithmetic on both sides with 2026 numbers: real token prices, real hardware tiers, worked crossover examples, and the honest cases where each side wins.

How LLM API Pricing Actually Works in 2026

Every major provider bills the same way: a rate per million input tokens and a higher rate per million output tokens. The 2026 price sheet (verified against vendor pages by independent trackers) looks like this:

• Frontier tier: GPT-5 at $1.25 input/$10 output per million tokens; Claude Sonnet 5 at $2/$10; Claude Opus 5 at $5/$25; Gemini 2.5 Pro at $1.25/$10.
• Small-model tier: GPT-5-mini at $0.25/$2; Gemini 2.5 Flash at $0.30/$2.50; Flash-Lite at $0.10/$0.40. OpenAI's cheapest current model floors near $0.05/$0.40.
• Discounts that change the math: both OpenAI and Anthropic offer a 50 percent batch discount on non-urgent jobs, and Anthropic's prompt caching reads repeated context at one-tenth of the standard input rate - so cache-friendly workloads cut the input bill by an order of magnitude before a single hardware dollar is spent.

The cost curve is linear, and at scale it gets steep. Red Hat's analysis of hidden LLM costs estimates that a startup handling 10,000 requests per day with a capable model can spend around $15,000 per month on API calls, and a Dell Technologies white paper pegs GPT-4o-class inference at roughly $12.19 per user per month for a 70B-class workload - which is why finance teams start asking what a GPU would cost instead.

What Self-Hosting Actually Costs (Hardware, Power, and Your Time)

Self-hosting swaps the per-token meter for three cost layers, and the first two are easy to understate:

• Hardware, one-time. A used RTX 3090 (24 GB VRAM, roughly $700-$900) runs 7B-14B models comfortably and 32B models compressed; an RTX 4090 build ($1,600-$2,000) does the same faster; two used 3090s (48 GB, about $1,600) reach 72B-class models compressed. A practical business rig for 14B-32B models at 4-bit quantization lands near $2,500, while serious multi-GPU servers run $30,000-$60,000 and beyond.
• Electricity. An RTX 4090 draws about 300 W under inference load. At the US-average $0.15 per kWh that is roughly $10.80 per month running 8 hours a day, or about $32 per month at 24/7 (EIA's 2025 retail average was 13.63 cents per kWh, so real bills vary by region). Power and cooling add another 15-25 percent to total cost of ownership at typical US rates, more in Europe (Presenc AI's 2026 TCO analysis).
• Your time. Model updates, driver compatibility, quantization tuning and uptime are your problem now. Budget a few hours a month of engineering time, and treat it as real cost - it is the line item most comparisons quietly omit.

After purchase, every token is electricity: a continuously running consumer GPU generates a million output tokens for a few cents of power - versus $10 for the same million output tokens on Claude Sonnet 5. That gap is the entire argument for self-hosting, and it is real.

The Crossover Math, Worked

Here is the cleanest worked example, built on published API rates. Suppose your workload matches a Llama 3 70B-class model served on Groq at $0.64 per million input and $0.80 per million output tokens. A typical request (1,000 input tokens, 500 output) costs about $0.001. A used RTX 3090 costs about $700:

• At 1,000 requests per day (30,000/month), the API bill is roughly $30 per month - the GPU pays for itself in about 23 months. Marginal.
• At 10,000 requests per day (300,000/month), the API bill is roughly $300 per month - payback drops to about 2.5 months. Compelling.
• Against frontier models the crossover arrives far sooner: Presenc AI's 2026 analysis finds 70B-class self-hosting reaches breakeven against frontier cloud APIs in 3-6 months at moderate utilization, and 7B-class workloads at 30 percent workstation utilization cross in 4-9 months. But at under 10 percent utilization - sporadic use - the breakeven horizon stretches to 2-4 years, and the API stays cheaper for years.

Team size moves the same curve from the seat side. ChatGPT Business runs roughly $25 per user per month, so the subscription bill scales linearly with headcount while a shared rig's cost is fixed:

• 3 people: about $900/year on Business vs a ~$2,500 rig - breakeven near 2.8 years, rarely worth it.
• 10 people: about $3,000/year - breakeven around 10 months.
• 30 people: about $9,000/year - breakeven in weeks to a few months. The crossover point is headcount, not enthusiasm.

One caution from careful 2026 modeling: the crossover can sit just past your hardware's capacity. One full-cost analysis found a lean $4,699 local system crossing current GPT-5.6-class pricing only at about 5,048 requests per day - against a single-node capacity of roughly 5,750. A crossover you can barely reach is not a business case, and adding a spare for redundancy pushes it out of reach entirely.

When Self-Hosting Wins

• Steady, high, batchable volume. Embeddings, classification, summarization and support-triage pipelines saturate a GPU and pay nothing per token. Moving one high-volume classification job to a small local model cut costs roughly 90 percent in a widely cited operator example - the shape of that win repeats wherever the task is repetitive.
• The split-workload pattern. The sensible 2026 architecture many teams land on: a small self-hosted model handles the boring, repeatable 80 percent of tasks, and a frontier API handles the hard 20 percent that genuinely needs it. The hardware bill is amortized by volume; the premium tokens are reserved for the work that deserves them.
• Data privacy and jurisdiction. When prompts contain client records, contracts or regulated data, cost stops being the deciding factor. If data-residency obligations or contractual promises mean prompts cannot leave your perimeter, self-hosting (or a private cloud under your control) is the only compliant shape - and the cost comparison becomes about which compliant option is cheaper, not whether compliance is worth it.
• Latency, offline operation and rate limits. No provider outages, no throttling, no per-seat gatekeeping - and for air-gapped or offline requirements there is no API at all.

For a full hardware walkthrough - VRAM math, build tiers from $2,000 to $40,000 - see our guide to on-premise AI servers for small business.

When the API Stays the Right Answer

• Bursty or low volume. Under 10 percent utilization, hardware idles and the API wins by a mile; a three-person team paying $900/year cannot amortize a $2,500 rig plus someone's maintenance time.
• Frontier-quality requirements. The hard constraint nobody prices around: you cannot self-host GPT-5 or Claude Opus 5 at all. If the work needs frontier reasoning, the comparison is over regardless of the arithmetic.
• Zero-ops teams. A product team shipping an AI feature should start on the API; the $50-$200/month bill during development is cheaper than the engineering time to run an inference stack, and model updates are someone else's 2 a.m. problem.
• The "90 percent cheaper" trap. Self-hosting does not make cost disappear - it changes shape: from a per-token meter you can watch to fixed hardware, power and your own time, paid whether you use it or not. Teams that model only the hardware price and skip power, redundancy and maintenance routinely discover the crossover they computed never arrives in practice.

Frequently Asked Questions

What does it cost per month to run a self-hosted LLM?

After hardware, mostly electricity: a single RTX 4090 at 300 W costs about $10.80/month at 8 hours a day or roughly $32/month running 24/7 at $0.15 per kWh. Add 15-25 percent for power and cooling overhead, plus a few hours of monthly maintenance time - the tokens themselves cost effectively nothing once the GPU is paid for.

At what usage level does self-hosting beat the API?

Against mid-tier APIs, a consumer GPU crosses over around 1,000+ requests per day within one to two years, and around 10,000 requests per day within months. Against frontier APIs, breakeven arrives in 3-6 months at moderate utilization because the per-token premium is 5-50x higher. Below 10 percent utilization, the breakeven horizon stretches to 2-4 years and the API remains cheaper.

Can I self-host GPT-5 or Claude?

No - those are proprietary, closed-weight models that exist only as hosted APIs. Self-hosting means open-weight models (Llama, Qwen, Mistral, Gemma, gpt-oss), which trail the frontier on complex reasoning but are genuinely strong for classification, extraction, summarization and routine drafting. If your workload demands frontier quality, the cost comparison is moot.

Is self-hosting worth it for a small team?

Usually not on pure cost: a 3-person team on ChatGPT Business pays about $900/year, which takes nearly 3 years to recover on a $2,500 rig. Ten people reach breakeven around 10 months, and 30 people in weeks to months. Small teams self-host for privacy or control reasons, not savings - and should budget the maintenance time honestly.

If you want a straight answer on where your workload's crossover sits - measured on your own request volumes, not someone else's benchmarks - contact ArcBeta for a private AI deployment assessment. We model the API bill against the hardware bill with your numbers and tell you honestly if the subscription should stay.