Is self-hosting an LLM cheaper than using an API?

It depends entirely on volume and which API you're comparing against. Self-hosted inference on rented GPUs runs roughly $0.15 to $0.80 per million tokens depending on model size, versus $2 to $15 per million tokens for frontier APIs, a real 10 to 20x gap at sustained high volume. But that gap only pays back the fixed GPU and engineering cost above a breakeven volume, and that breakeven volume moves by more than an order of magnitude depending on which API you'd otherwise be calling.

What is the break-even point for self-hosting an LLM vs calling an API?

There is no single number. Priced against an expensive flagship model like GPT-5-class pricing, self-hosting a comparable open-weight model breaks even around 256 million tokens a month, roughly 8.5 million tokens a day. Priced against a cheap, high-throughput model like DeepSeek V4 Flash or Gemini 2.5 Flash-Lite, the breakeven volume runs into the billions of tokens a month, a range most teams will never reach. A commonly cited general rule puts self-hosting an open-weight model at 'economically rational' somewhere between 50 and 200 million tokens a month, but that range assumes you're displacing a mid-to-expensive API, not the cheapest one available.

How much does it actually cost to self-host an LLM per month?

For a single A100 80GB running 24/7 at roughly $2/hour, the raw GPU bill is about $1,440 a month; a 2xA100 setup runs closer to $2,880. Add engineering time, 10 to 20 hours a month of maintenance and monitoring at $75 to $150 an hour for a senior DevOps or ML engineer, and that's another $750 to $3,000 a month. A common rule of thumb multiplies the raw GPU cost by 1.3x to 2x to get a realistic total cost of ownership, which puts a single well-run A100 deployment around $3,000 to $3,500 a month before it has served a single production token.

Does engineering time really change the self-host vs API math?

Yes, and it's the line item most cost comparisons skip. A maintained self-hosted deployment needs 10 to 20 hours a month of a senior engineer's time for patching, monitoring, and troubleshooting, which prices out to $750 to $3,000 a month at typical DevOps or ML engineer rates. On a small deployment, that labor cost can exceed the GPU rental cost outright. Any self-hosting cost comparison that only counts GPU-hours and skips engineering time is undercounting the real bill by a meaningful margin.

Should I self-host, use a managed open-weight API, or call a frontier API?

Below tens of millions of tokens a month, a managed API, either a hosted open-weight provider or a frontier lab, almost always wins once engineering time is counted, because there's no fixed cost to amortize. Self-hosting starts to make sense once volume is large and steady enough to keep GPUs busy most of the time, and the model you'd otherwise call is expensive enough that the per-token gap is worth the fixed cost, which is why the breakeven volume against a frontier flagship is so much lower than the breakeven volume against a cheap open-weight model. Most teams don't need to pick once and commit; a routing layer can send volume to whichever option is cheapest for a given request as pricing and volume both shift.

The Real Monthly Bill

Self-hosting an LLM breaks even around 256 million tokens a month against a flagship frontier model, but closer to 6.9 billion against DeepSeek V4 Flash. There are three ways to run an LLM in production, and they carry three completely different bills: self-host a GPU fleet, call a managed API for an open-weight model, or call a frontier API. The self-host line looks free right up until you price the GPUs, the on-call rotation, and the 10 to 20 engineer-hours a month it takes to keep vLLM patched. The breakeven volume where self-hosting starts to win is not one number either. Same GPUs, same team, a 27x difference in the answer, because the number that actually moves is the price of whatever you're not calling anymore.

Three ways to run a model. Three completely different bills.

There are three ways to put an LLM behind your product, and teams keep comparing them like they're the same purchase with different price tags. They aren't. Self-hosting is buying fixed capacity. A managed API for an open-weight model is buying variable, pooled capacity at a thin margin. A frontier API is buying the model itself, priced at whatever the lab that trained it decides to charge. Each one carries a real monthly number, and the "free" one on that list is very often not the cheapest.

What each option is actually made of

	Self-host	Managed open-weight API	Frontier API
What you're paying for	Rented or owned GPUs, 24/7	Someone else's GPUs, pooled across customers	The model itself, plus the lab's own compute
Cost shape	Fixed: the bill is close to the same whether you send 1 token or 10 million	Variable: scales with tokens sent	Variable: scales with tokens sent
Who absorbs utilization risk	You. Idle GPU-hours are still billed hours	The provider, pooled across every customer on the platform	The lab, pooled across every customer on the platform
Engineering time	Real and recurring: patching, monitoring, scaling	Near zero	Near zero
Typical $/M tokens	$0.15-$0.80 depending on model size	Roughly $0.05-$0.90 depending on model and provider	$2-$15 depending on model and lab
Where it wins	High, steady volume against an expensive alternative	Low-to-mid volume, open-weight quality is enough	Any volume where the quality ceiling matters more than the per-token price

The middle column is the one most "self-host vs API" comparisons skip entirely, which is a real gap, because a managed open-weight API gets you open-weight pricing without taking on any of the fixed cost or utilization risk in the left column.

The GPU math, priced honestly

Cloud H100 rental runs $1.49 to $6.98 an hour across more than a dozen providers, averaging around $3.20 to $3.39, with hyperscaler on-demand pricing running well above that: AWS around $6.88/hour, Azure around $12.29/hour, versus specialized GPU clouds in the $2-3/hour range. Source: IntuitionLabs, "H100 Rental Prices Compared," 2026; Source: Spheron, "GPU Cloud Pricing Comparison 2026," 2026. A single A100 80GB at roughly $2/hour runs about $1,440 a month rented around the clock; two of them run about $2,880. Source: DevTk.AI, "Self-Host LLM vs API: Real Cost Breakdown 2026".

That's before the fudge factor every honest TCO estimate adds. A common rule of thumb multiplies the raw GPU rental cost by 1.3x to 2x to reach a realistic total cost of ownership, once load balancing, monitoring, and redundancy are priced in, which puts a single well-run A100 deployment around $3,000 to $3,500 a month before it has served a single production token. Source: DevTk.AI, "Self-Host LLM vs API: Real Cost Breakdown 2026". Fixed monthly GPU spend for a Llama-3-class 8B model on an L40S lands around $900 to $1,500; a 70B model on A100s runs $1,500 to $2,500. Source: Tensoria, "Self-Host LLMs or Use the API? Real $/1M Token Numbers in 2026".

None of that includes the part almost every calculator leaves out.

The line item that gets skipped: your own time

A maintained self-hosted deployment needs 10 to 20 hours a month of a senior engineer's time for patching vLLM, watching dashboards, and chasing down the 2am page. At $75 to $150 an hour for a senior DevOps or ML engineer, that's $750 to $3,000 a month, recurring, whether or not the fleet is busy. Source: DevTk.AI, "Self-Host LLM vs API: Real Cost Breakdown 2026". On a single-GPU deployment, that labor line can be larger than the GPU rental itself. It's also the part of the bill nobody puts on a pricing page, which is exactly why it's the part that gets left out of the comparison.

Add the GPU rental, the TCO multiplier, and the engineering time together and a small self-hosted deployment, the kind a mid-sized team would actually run, lands in the same $5,000 to $15,000-a-month range that shows up whenever anyone actually itemizes it rather than quoting the GPU-hour price alone.

The real monthly bill by hosting model: self-host (2xGPU plus ops) runs $5k to $15k a month, a managed open-weight API runs $200 to $2k a month, and a frontier API runs $1k to $4k a month, for a representative mid-volume deployment

The breakeven volume is not one number

Here's the finding that should change how this comparison gets made. Priced against different APIs, the volume where self-hosting starts to win moves by more than an order of magnitude:

Against a flagship frontier model, self-hosting a comparable open-weight model breaks even around 256 million tokens a month, roughly 8.5 million tokens a day.
Against DeepSeek V4 Flash, a fast, cheap model near the bottom of the market, the breakeven volume runs to roughly 6.86 billion tokens a month.
Against Gemini 2.5 Flash-Lite, similarly cheap, breakeven runs to roughly 5.76 billion tokens a month.

Source: DevTk.AI, "Self-Host LLM vs API: Real Cost Breakdown 2026". A separate, more general estimate puts self-hosting an open-weight model at "economically rational" somewhere between 50 and 200 million tokens a month, which sits in between those two extremes because it assumes you're displacing a mid-priced API, not the cheapest one on the market and not the most expensive. Source: Tensoria, "Self-Host LLMs or Use the API? Real $/1M Token Numbers in 2026".

The mechanism is simple once you see it: breakeven volume is the fixed self-hosting cost divided by the per-token price gap. Compare against an expensive model and the gap is wide, so it takes fewer tokens to pay back the fixed cost. Compare against a model that's already cheap, and the gap nearly closes, so the fixed cost never gets paid back at any volume a normal product will reach. The self-host-versus-API decision was never really about self-hosting. It's a bet on how expensive the alternative stays.

Where the three cost lines cross as monthly token volume grows: self-host is a flat fixed cost, frontier API cost overtakes it around 20 million tokens a month, and managed open-weight API cost overtakes it around 50 million tokens a month, illustrative slopes

And the fixed line only holds if you keep the GPUs busy

Every number above assumes the self-hosted fleet runs close to saturated. That assumption is doing a lot of work. A June 2026 paper found that identical H100 hardware costs anywhere from $0.21 to $15.25 per million output tokens purely as a function of GPU utilization, a 2.5 to 24x penalty at ordinary 1-10 requests-per-second business load and up to 36.3x near idle. A fleet sized for peak traffic runs underutilized most of every day it operates, which means the flat line in the chart above is really a floor, not a guarantee. Get the utilization wrong and the breakeven volume you calculated on paper is the volume you'd need at a fully saturated fleet, not the one you'll actually see.

A decision framework, not a rule of thumb

Under roughly 20 to 50 million tokens a month: a managed API almost always wins once engineering time is counted, because there's no fixed cost to amortize and no utilization risk to carry.
Well above that, against an expensive model: self-hosting can pay back its fixed cost quickly, sometimes within tens of millions of tokens a month, if the fleet stays busy.
Against a cheap, already-efficient model: don't bother. The per-token gap is too thin to ever clear a GPU rental bill plus an engineer's time, and routing to that cheap model directly captures most of the same saving with none of the fixed cost.
Managed open-weight APIs are the option most comparisons skip: they get you open-weight pricing and open-weight model choice without the fixed cost or the utilization bet, and they're very often the right first move before a self-hosting decision is even on the table.
Data residency and compliance can override all of this. If classification or generation has to stay inside your network boundary, that's a real constraint self-hosting solves that no API pricing table changes.

Why this doesn't have to be a one-time bet

The uncomfortable part of the analysis above is that the right answer moves. DeepSeek V4 Flash and Gemini 2.5 Flash-Lite didn't exist at their current prices a year ago, and the next round of aggressive open-weight pricing will move the breakeven line again, probably in the same direction: further out of self-hosting's reach for anything but the highest, steadiest volumes. A team that built a GPU cluster to beat last year's frontier pricing is now competing against this year's cheap model, on infrastructure they already paid for.

Nadir sits underneath that decision instead of locking you into one side of it. Point your OpenAI-compatible calls at Nadir with model=auto and every request routes to whichever model, hosted frontier, hosted open-weight, or your own self-hosted fleet, is the cheapest one that can actually answer it right now. If your volume is large and steady enough that self-hosting some of the traffic makes sense, Nadir routes across your own infrastructure too, so the fixed capacity you already built gets the requests worth sending to it while the rest goes wherever the market has moved. You don't have to re-run this analysis every time a lab cuts a price.

Bottom line

Self-hosting isn't free, and it isn't a fixed win at scale either. Priced honestly, a small self-hosted deployment runs $5,000 to $15,000 a month once GPUs, redundancy, and engineering time are counted, and the volume where that fixed cost pays for itself depends entirely on which model you'd otherwise be calling, from 8.5 million tokens a day against a frontier flagship to a volume most products will never reach against the cheapest models on the market today. Run the math against your actual alternative, not the one that was true last year, and keep the utilization number honest. If you'd rather not re-run it every time pricing shifts, that's the layer routing is built to own.

Start free, or read the utilization math that decides whether your self-hosted fixed cost actually holds.

Data in the charts is illustrative, modeled from the cited research to show the shape of the tradeoff, not a live pricing feed. Sources: [IntuitionLabs, "H100 Rental Prices Compared: $1.49-$6.98/hr Across 15+ Cloud Providers," 2026](https://intuitionlabs.ai/articles/h100-rental-prices-cloud-comparison). [Spheron, "GPU Cloud Pricing Comparison 2026," 2026](https://www.spheron.network/blog/gpu-cloud-pricing-comparison-2026/). [DevTk.AI, "Self-Host LLM vs API: Real Cost Breakdown 2026"](https://devtk.ai/en/blog/self-hosting-llm-vs-api-cost-2026/). [Tensoria, "Self-Host LLMs or Use the API? Real $/1M Token Numbers in 2026"](https://tensoria.fr/en/blog/deploying-llms-to-production). [Patil, "Beyond Per-Token Pricing: A Concurrency-Aware Methodology for LLM Infrastructure Cost Estimation," arXiv:2606.11690, June 2026](https://arxiv.org/abs/2606.11690).

More on pricing & models

What Nadir is

Nadir is an LLM router. Nadir sends each prompt to the cheapest model that can handle it. A small decision model reads each prompt against your model menu and picks the lowest-cost model it expects to handle it. When the decision model is uncertain or unavailable, the request goes to the strongest model on the menu. The routing step makes no call to your model provider.

Nadir runs two ways. The decision API returns a model, reasoning-effort, cache, context, and policy recommendation without calling a model provider, beside the gateway you already run. That is how a shadow-mode evaluation works, and its projected savings stay advisory. The OpenAI compatible managed proxy executes the route, and migration is a two-line change: point the base URL at api.getnadir.com and set model to auto. On that path an optional verifier can score a complete non-streaming answer and escalate to a stronger model when it misses the configured bar. Streaming bypasses post-generation verification. BYOK is supported on every tier.

For coding agents, Nadir Auto is a terminal launcher for Claude Code or Codex that asks Nadir for the main-session model before each turn, with your configured model as the fallback and cost ceiling. Delegation integrations for Claude Code, Codex, or Cursor instead recommend a model tier for subagent work, and the agent decides; the Claude plugin can also move a main-thread turn to Haiku when Nadir's cache check prices it cheaper, unless its switch_main_model option is off. Either way inference runs on your own provider account, so Nadir holds no provider keys on that path.

What the numbers are, and what they are not

Nadir publishes each evaluation with its scope. On checkable code, run-check-escalate solved 392 of 395 common HumanEval and MBPP problems (99.2%), graded by running the canonical tests; that applies only to tasks with runnable deterministic tests. Nadir-Tumbler posts an arena_score of 72.3 on RouterArena's public scorer, 5th of 23 routers, which measures the routing decision on RouterArena's own model pool. A reference-assisted RouterBench evaluation over 11,420 held-out triples produced a 60% lower projected cost than always-Opus with about 98% retained quality and a 1.7% catastrophic-route rate. That experiment gave the verifier the expensive-model reference answer, which production does not have, so it is a research ceiling and not the deployed path.

None of these is a production guarantee, a universal savings rate, or a forecast for any particular workload. Customer savings are reported from measured execution against a declared baseline, and customer quality only from outcome-labelled traffic. Projected savings and realized savings are separate artifacts and are never blended.

Design-partner program

Three rungs, picked by risk appetite. Rung 0 Shadow runs advisory decision calls alongside live traffic and returns a projected receipt, with nothing in the request path changed. Rung 1 Hosted is the two-line swap on a production slice and returns a realized receipt. Rung 2 On-prem is a supervised six-week proof of concept inside the partner's VPC, where no prompt, response, or usage reaches Nadir. The commitments are the same at every rung. Apply for a rung directly: Rung 0 Shadow, Rung 1 Hosted, or Rung 2 On-prem. Not sure which fits? Start at getnadir.com/contact/?reason=design-partner.

Licensing

NadirClaw is the self-hosted core, source-available under the PolyForm Noncommercial License. Source-available is the correct label; NadirClaw is not open source. Nadir Route's hosted plan has no base fee and charges a variable fee only on measured savings from requests Nadir executed.