Replicate Per-Second Rates, Cold Starts & the Real Bill 2026 Guide
Replicate charges by the second, so an idle account is free and a busy one is not. Hardware sets the rate, cold starts bill before any work, and the big rigs need a contract. Here is the real math.
Typical compute rate
$0.09-$5.49/hr
CPU to H100, billed per second; no monthly subscription
Hidden fees
Yes
cold starts billed, idle warm time charged, multi-GPU contract-gated
Free tier
Pay-as-you-go
no monthly floor; an idle account costs nothing
Cost transparency
Medium
scores 4 of 6 on our transparency checklist
Replicate true cost, by the second
High· Verified July 15, 2026Replicate really costs nothing at rest and bills per second while a model runs as of July 15, 2026, so there is no monthly floor. Hardware sets the rate. A CPU is $0.09 an hour, a T4 $0.81, an A100 80GB $5.04 and an H100 $5.49. An A100 kept warm for a day runs near $121. Cold starts bill before inference begins, and multi-GPU rigs like the $43.92-an-hour 8x H100 are contract-gated. Enterprise adds volume discounts, so committed spend is the only lever on the rate.
- CPU /hr$0.09
- T4 GPU /hr$0.81
- A100 80GB /hr$5.04
- H100 /hr$5.49
- A100 warm, per day$121
- 8x H100 /hr$43.92
- Monthly floor$0
Replicate carries no monthly seat, so it sits outside the $7.99 median across the 20 llm tools we track. An idle account is free, but an A100 kept warm runs about $121 a day.
The free side of Replicate, and its limits
There is no free plan, but there is something better for light use: no monthly floor at all. An idle Replicate account costs $0, because you pay only while a model actually runs. That makes it a natural fit for spiky, occasional or experimental workloads, where a subscription would charge you for time you never used.
The catch arrives with scale. Per-second billing has no ceiling, so steady production traffic tracks demand exactly, and cold starts add cost every time a scaled-down model wakes. Prototype freely, then watch the meter as traffic grows. If your workload turns steady and heavy, weigh running it here against a managed API or your own hardware on the Replicate alternatives page. Per-second compute is not always the cheapest home for a busy model.
Replicate savings that come from how you deploy
Replicate has no seats to discount and no coupons on offer. It runs no student or nonprofit program as of July 2026, for the reason most usage platforms share: a usage bill leaves a coupon nothing to discount. The savings all live in deployment choices, and they are the ones that actually move a per-second bill.
Pick the smallest hardware that meets your latency bar, tune concurrency so a warm instance serves more per second, and decide deliberately between scale-to-zero and keeping capacity warm. For steady, high-volume traffic, the Enterprise tier offers volume discounts on compute, and that quote-based lane is where the negotiation section does its work.
Pick the smallest chip that fits
Moving a model off an A100 at $5.04 an hour onto a T4 at $0.81, where it fits, cuts the rate six-fold. Matching the chip to the model is the cleanest saving on the whole platform.
Tune the cold-start tradeoff
Scale-to-zero saves money on a rarely-hit endpoint but pays a cold start on each wake. Keeping capacity warm avoids that but bills continuously. The right choice depends on your traffic shape, and getting it right saves either way.
Enterprise volume discounts
Steady, high-volume compute qualifies for Enterprise volume pricing below the public per-second card, with SOC 2, private deployment and an SLA. It is quote-based, so committed spend is the lever that opens the discount.
Batch and consolidate runs
Packing more work into each warm window, rather than spinning hardware up and down repeatedly, spreads the cold-start cost across more predictions. For bulk jobs, consolidating runs meaningfully lowers the effective per-prediction rate.
Negotiating a Replicate Enterprise agreement
The per-second card is fixed. Nobody discounts a single T4 or A100 hour, and the self-serve dashboard is the same rate for everyone. Negotiation opens at Enterprise, where volume discounts, private deployment and an SLA are quote-based, and your committed compute spend is what moves the number.
Three plays cover most of the ground. Each turns on the same point: Replicate prizes a committed compute floor over spot traffic on the meter.
Commit compute for a lower per-second rate
- Target
- Enterprise volume
- Argument
- Guarantee a monthly GPU-hour spend in exchange for a rate below the public $0.09-to-$5.49 card. Forecastable compute revenue is worth a discount, and a predictable bill is worth it to your team.
Price the workload against a managed API
- Target
- Any inference-heavy contract
- Argument
- If the job is standard inference, Amazon Nova at $0.035 per million skips compute management entirely. Make Replicate justify the per-second model, or match the economics on your real traffic.
Fold in an SLA and private deployment
- Target
- Enterprise renewal
- Argument
- If you are committing spend, get the latency SLA, SOC 2 terms and VPC peering in the same contract. Reserved capacity without those guarantees is just a higher bill, so bind them to the rate.
When timing a Replicate deal actually helps
The per-second meter has no timing angle. A GPU-hour costs the same in July as in December, so the calendar only bears on an Enterprise commitment. Replicate's sellers run quarterly targets, so a committed-compute deal settled in a quarter's closing days usually reads better than one begun at its opening.
Jan
Feb
Mar
Q-END
Apr
May
Jun
Q-END
Jul
Aug
Sep
Q-END
Oct
Nov
Dec
Q-END
Pro tip: Come with an audit of your real compute usage, not a projection you hope to hit. A credible GPU-hour number is what earns a rate below the public card, and a vague growth story is not.
Replicate costs: what flexes and what does not
Point requests where Replicate can actually move. The per-second card holds for every self-serve user; negotiating room lives in committed compute and Enterprise terms.
Usually negotiable
- Committed per-second compute rateHIGH
- Multi-GPU and reserved capacity pricingHIGH
- Latency SLA and private deploymentMEDIUM
- SOC 2 and VPC peering termsMEDIUM
- Payment terms (Net 30/60)LOW
Rarely negotiable
- Published per-second hardware rates
- Per-token rates on token-billed models
- Cold-start billing behavior
- The self-serve single-GPU tiers
Replicate negotiation email generator
Add your details and the message below drafts itself, and it fills in rival compute and token rates live from our verified catalog. Route the finished note to your Replicate Enterprise contact, or drop it in the sales form. Keep it lean. Give your monthly GPU-hour spend, put a competing rate next to it, ask for a committed per-second rate plus an SLA, and pin the date work can start.
committed compute, SOC 2, private deployment, SLA
Hi Replicate team, I lead tooling decisions at [Your company], and we are evaluating an enterprise credit pool for our team of 10-50 people. As part of this evaluation we are also looking at Hugging Face, which comes in at $9/mo + metered GPU, and Amazon Nova at $0.035 per 1M input. Can you help us understand the value difference at your current rates? We are ready to commit to an annual term. What is the best rate you can offer on annual billing, and can you cap the renewal price in the contract? We are aiming to sign before the end of this quarter, and budget sign-off is already in place. Could you share a proposal covering the per-seat or per-credit rate, the renewal terms, and any programs we qualify for? Best regards, [Your name] [Your company]
Send it Tuesday to Thursday, and follow up once after 3 business days.
Before you send
- Have your real GPU-hour spend ready. A volume ask with no compute number behind it goes nowhere.
- Send it midweek, since a note arriving Tuesday through Thursday clears the queue faster than Monday or Friday.
- Do not lead with your budget. Let Replicate put the first committed rate on the table.
- Name two compute rivals in the note. The generator fills their current rates in for you.
- Get the per-second rate, the SLA and the deployment terms in writing before you shift production.
- Follow up once after a few business days, then read continued quiet as an answer of its own.
Replicate billing mistakes that drain compute budget
Each of these springs from how per-second billing works, and every one is avoidable before you deploy.
Keeping an A100 warm around the clock. At $5.04 an hour, that is about $121 a day whether or not it runs.
Oversizing the hardware. A model that fits on a T4 at $0.81 has no reason to sit on a $5.04 A100.
Ignoring cold-start cost on rare endpoints. Each wake bills at the hardware rate before any inference happens.
Assuming a model is time-billed. Some price per token instead, so check before you estimate the run cost.
Planning around multi-GPU on demand. The 8x rigs are contract-gated, so budget them through Enterprise, not the card.
Replicate rivals worth benchmarking on cost
On a per-second platform, leverage is a spreadsheet, not a speech. These three give you a real point of comparison for running a model, verified in our catalog. You do not have to move. You need a benchmarked number from one of them, so a committed-rate conversation with Replicate rests on evidence rather than a bluff.
Hugging Face
PRO seat plus metered Spaces GPUs
$9/mo
The other run-any-model platform. Its Spaces GPUs meter much like Replicate's, so it is the direct cross-shop when the question is where to host your own model.
Amazon Nova
managed inference, no compute to run
$0.035/1M
The managed-API floor. If your workload is standard inference, Nova removes the per-second compute entirely, which is the sharpest angle against renting GPUs.
Mistral Large
input, batch 50% off, EU-hosted
$2/1M
A managed frontier model with European hosting. The alternative when the pull to self-host on Replicate is really about control rather than raw price.
Script“For steady inference, Amazon Nova is $0.035 per million and Hugging Face runs metered Spaces GPUs. What committed per-second rate can Replicate offer against those?”
Is Replicate worth it? A per-second cost read
Replicate is the right tool for spiky and experimental work and the one to watch for steady heavy load, and the per-second model explains both. An idle account is free, which is genuinely hard to beat for occasional jobs. A busy one is uncapped, and cold starts and warm idle time add cost that never appears as a line on a plan card.
So manage the meter. Right-size the hardware, tune the scale-to-zero tradeoff to your traffic, and consolidate runs to spread cold-start cost. Watch the bill as volume grows, and price a managed API or your own hardware against it. Once your compute is steady and real, take an Enterprise conversation to committed spend for a rate under the card.
Do that and Replicate stays efficient across a wide range of workloads. The full hardware and model rates live on the Replicate pricing page. This guide exists to keep the per-second meter from outrunning the value.
Replicate pricing and discount FAQ
How does Replicate pricing work?
+
Replicate has no subscription. It bills per second of compute while a model runs, and the rate is set by the hardware. A CPU is $0.09 an hour, a T4 $0.81, an A100 80GB $5.04 and an H100 $5.49. Some hosted models bill by token instead, around $0.10 per million input and $0.50 output. An idle account costs nothing because you pay only for active run time, but there is no monthly cap, so a busy production workload tracks demand exactly.
Why am I charged when my Replicate model is idle?
+
Two reasons. If you keep a model warm to avoid cold starts, it bills for every second it stays up, traffic or not, which is about $121 a day on an A100. If you use scale-to-zero, the model spins down when idle but pays a cold start when the next request wakes it, billed at the hardware rate before inference begins. Neither is a bug. Both are the tradeoff of per-second compute, and choosing between them is the main way to control idle cost.
Does Replicate cost anything when idle?
+
There is no paid subscription and no fixed free plan, but an idle account costs nothing because billing is purely per second of active compute. That makes it effectively free to keep around for occasional or experimental use. You start paying the moment a model runs, at the hardware rate, and there is no monthly ceiling. For light, spiky workloads that is cheaper than a subscription; for steady heavy traffic, the uncapped meter is the thing to watch.
Are there discounts or credits on Replicate?
+
Replicate publishes no academic, nonprofit or startup rate as of July 2026, because a usage bill gives a coupon nothing to reduce. The savings that matter are in how you deploy: right-sizing the hardware, tuning the cold-start tradeoff, and consolidating runs. For steady, high-volume compute, the Enterprise tier carries volume discounts below the public per-second card, along with SOC 2, private deployment and an SLA. That quote-based lane is where real discounts are negotiated.
How much does an A100 cost on Replicate?
+
An A100 80GB is $5.04 an hour, billed per second. Kept warm for a full day that is roughly $121, whether it serves one prediction or thousands. An H100 is a little higher at $5.49 an hour. If you need parallel capacity, the multi-GPU rigs are contract-gated: an 8x H100 lists at $43.92 an hour but is not available on demand. Match the chip to the model, since a workload that fits a $0.81 T4 has no reason to run on an A100.
How do you get a volume discount on Replicate?
+
Through the Enterprise tier. The per-second card is fixed for self-serve users, but Enterprise offers volume discounts on compute, private deployment and an SLA, all quote-based. Commit a monthly GPU-hour spend for a rate below the public card. Bring a managed-API benchmark like Amazon Nova at $0.035 per million to anchor the conversation. Fold the SLA and SOC 2 terms into the same contract. Reserved compute is worth a discount, so expect 10 to 25 percent at genuine volume.
Should I use Replicate or a managed inference API?
+
For custom or open models you want to run yourself, Replicate is flexible and often cost-effective. For standard inference, a managed API like Amazon Nova at $0.035 per million tokens can be cheaper. It carries no per-second compute, no cold starts and no idle warm time. The honest test is to benchmark your real traffic on both. Self-hosting on Replicate wins on control and model choice; a managed API often wins on pure cost for steady, common workloads.
What is the cheapest way to run a model on Replicate?
+
Match the hardware to the model and control idle time. Run a model on the smallest chip it fits, a $0.81 T4 rather than a $5.04 A100 where possible. Decide deliberately between scale-to-zero and keeping capacity warm, based on your traffic. Consolidate bulk work into warm windows to spread cold-start cost across more predictions. For a steady, heavy workload, price a managed API against the per-second bill. Those habits keep the compute meter honest.
Explore Replicate
Every page on Replicate in one place, you are on cost guide.
Snapshot, score and verdict
How to get API access, limits, SDKs and what it costs
Latency, throughput, uptime and behaviour under scale
Every tier and the entry price
You are here
Compared and ranked vs peers
Price and feature change history
Browse the full Large Language Models category
Sources & verification
| Source | What was checked | Last checked |
|---|---|---|
| Replicate official pricing | Verified plan prices, renewal rates and credit allowances | July 15, 2026 |
| Replicate website | Official vendor website | July 15, 2026 |
| Replicate pricing on ComparEdge | Current prices for every plan, with the cost calculator | July 15, 2026 |
Every fact on this Replicate pricing page is tied to a named source and a verification date. Freshness-sensitive figures trace to the sources above; verify against the vendor before relying on them.