Founding seats now open

Unlimited AI tokens for a flat monthly fee. For people who run agents all day.

Frontier-class open models on our own GPUs, through the OpenAI-compatible API you already use. No volume cap, no cut-offs. And when the GPUs are full, you decide what happens: wait, take a smaller model, fail fast, or jump the queue. Never a silent downgrade.

Confirmation goes to .

Zero

token caps, overage bills, or surprise cut-offs.

One

flat price you can put on a card today.

Why we can promise unlimited when everyone else is walking it back

Most "unlimited" AI plans resell someone else's API. Every token costs them money. So the heaviest users get throttled, rate-limited into uselessness, or quietly banned.

We run our own GPUs. A GPU bills the same whether it is 10% busy or 90% busy. Your heavy usage does not cost us more. It fills hardware we already pay for. Our enemy is idle silicon, not you.

01 — Fixed cost

Our bill is GPU-hours, not your tokens

Marginal cost of your next request is effectively zero. That is the whole reason a flat fee holds up.

02 — Rate limits, not caps

At peak you may wait. You will never be metered.

We limit requests per second and concurrent sessions so one user cannot hog the queue. We do not count volume at all.

03 — Zero migration

OpenAI- and Anthropic-compatible endpoints

Keep your SDK, your agent framework, your eval harness. Change the base URL and the key.

04 — Your data stays yours

We do not log your prompts. We do not sell your data.

Requests are processed and discarded. No prompt or completion logging, no training on your traffic, no data brokers. We keep request counts for rate limiting, nothing more.

When we are busy

You choose what happens when the GPUs are full. We tell you which model answered. Every time.

Every other "unlimited" plan handles peak load the same way. It quietly swaps in a cheaper model and hopes you do not notice. We cannot do that, because the policy that fires is one you picked.

Requests queue first. If the wait passes your tier's deadline, your overflow policy decides. Set it once on your account, or per request with one header.

queue

Wait for the real thing

Hold your place until the model you asked for is free. You get a queued event with position and ETA, not a silent stall. Right for batch jobs.

degrade

Take a smaller model now

A fast answer from a smaller model, with the substitution disclosed in the response. Right for chat and quick questions.

fail

Fail loudly, retry on your terms

A 429 with an honest Retry-After computed from the live queue. Right for agent loops, where a mid-task downgrade can corrupt state.

priority

Never shed, never degraded

Reserved headroom. The frontier model answers, now. This is the Priority tier's default. It is what you pay for.

Model attribution on every response

The model that actually served your request is in the response metadata and your dashboard. No guessing.

Published routing policy

The rules that pick a model for each request are public, versioned, and change-logged. If a rule is not written down, we do not apply it.

No-silent-downgrade guarantee

If a downgrade happens, we flag it in the response and you can re-run the request at priority. An undisclosed downgrade is a bug, and we owe you for it.

Per request, when your orchestrator knows better than the default:

X-Overflow-Policy: fail       # critical agent step
X-Overflow-Policy: degrade    # summarization step, same workflow

Built for people who burn through tokens. Not for everyone.

This is for you if

  • You run agents that loop for hours and your API bill is the first thing you check in the morning
  • You run evals, sweeps, or research jobs and want to stop rationing them
  • You are an indie hacker or solo founder shipping AI products on a card, not a procurement cycle
  • You are a small AI-native team that would rather pay a flat fee than build your own inference stack

Probably not for you if

  • You need SOC 2, ISO 27001, or US/EU data residency on day one
  • You need a contractual latency SLA with teeth before you send a single request
  • Your finance team needs per-token usage reports for internal chargeback
  • You run latency-critical trading or anything where a queued request is a lost dollar

We would rather tell you this now than lose your trust after you sign up.

Founding pricing · price locked for 24 months from launch

You pay for access and position. Never for volume.

Every tier is unlimited. The difference is where you sit in the queue when the GPUs are busy, and which model answers. Founding seats keep these prices for 24 months from launch.

Standard

$20 /mo

or $17/mo billed annually

  • Unlimited tokens
  • Full API access, usage dashboard, model attribution
  • At peak: queue, then a smaller model by default. Switch to fail or queue per account or per request

For latency-sensitive agents

Priority

$79 /mo

monthly only, cancel any time

  • Everything in Standard
  • Your overflow policy is always priority
  • Bounded p99 latency, published monthly
  • At peak: never shed, never degraded. The frontier model answers, now

Batch

$9 /mo

or $6/mo billed annually

  • Unlimited tokens
  • Built for evals, sweeps, and overnight pipelines
  • Cheapest way to use the hardware
  • At peak: jobs wait for an off-peak window and may be preempted
Reserve a founding seat

Pick your tier in the reservation form. Switch tiers any time before launch. Annual billing on Standard and Batch only.

Founding seats

We buy GPUs against reservations, not signups.

A waitlist email tells us you are curious. A refundable deposit tells us how much hardware to provision. That is why founding seats go first. They lock in the founding price and hold a reserved place in the queue on launch day.

Deposit: your first month ($9, $20 or $79), credited at launch. Fully refundable any time before we go live, no questions asked.

When
First node targeted for early 2027. We go live once reservations fund it. We will tell you the exact date 14 days ahead.
Who
Founding seats first, then the free list in order. No interview, no approval step.
Updates
One email a month on hardware, models and the launch date. Replies go to a founder.
Refund
Any time before launch, in full, automatically if we never launch.
Tier you are reserving

Done

We will write to .

FAQ

Questions you should be asking

Not covered here? Email hello@getunmetered.ai. Replies come from a founder.

So what is the catch?

Peak hours. When the pool is full, your request queues first. If the wait passes your tier's deadline, the overflow policy you chose fires. You keep waiting, take a smaller model, or fail with a Retry-After. We disclose whatever happens in the response. Priority never overflows. Nobody is ever billed for volume.

What happens to my agent mid-loop if you are busy?

Set your policy to fail and you get a clean 429 with a Retry-After computed from the live queue. Your orchestrator retries instead of continuing on a quietly different model. If you would rather it never happen, that is what Priority is for.

Which models?

Four open-weight models, each with a job. Qwen3.8-27B for low-latency interactive traffic. DeepSeek-V4.1-Flash for high-volume work and the Priority tier. GLM-5.3 for agentic coding. Nemotron 3.5 Bolt for batch jobs and overflow. The lineup will change as better open models ship. The price will not.

Where does it run?

Our first site is in Paraguay, on some of the cheapest hydroelectric power in the hemisphere. That cost floor is a big part of how the flat fee stays flat.

Will my 24/7 agent get me banned?

No. Around-the-clock usage is exactly what the hardware is for. The only limits are requests per second and concurrent sessions, published per tier before you pay.

What happens to my deposit if you never launch?

You get it back in full. Deposits are refundable at any point before we go live. If we never launch, we refund them automatically.

What do you do with my prompts?

Nothing. We do not log prompts or completions. We do not train on your traffic. We never sell or share your data. The only thing we store per request is a count, for rate limiting. We run our own hardware, so no upstream provider sees your traffic.

Do I need to change my code?

Point your OpenAI or Anthropic SDK at our base URL and swap the API key. Streaming, tool calling, and JSON mode work as you expect.

Stop thinking about tokens.

Founding seats open in order of reservation. Capacity goes live when the first node is funded.

Reserve a founding seat