Novita AI Pricing: Models & Costs (2026 Guide)
Nov 10, 202615 min read
There is no subscription to pay. Novita AI is pay-as-you-go, so you can wire it into a workflow built in a tool like Activepieces without committing to a plan, and it runs three meters:
- tokens for model APIs
- hours for GPU instances
- seconds for serverless GPUs
Which meter you pick sets the bill more than any single rate does.
Novita AI pricing rates
What Novita AI is and how it bills
Novita AI is a cloud platform that sells hosted model APIs and rentable GPU compute from the same account, and the unit metered differs by product.
As of 2026-11-09, the table below shows each product's meter and what sits behind the price. You can see which lines charge only for work done and which keep charging while idle.
| Product | Meter | Price | Inclusions | Limits |
|---|---|---|---|---|
| Token API (Llama 3.3 70B, DeepSeek V3 0324) | Per million input and output tokens | Set per model, with separate input and output rates | Hosted inference, no server to manage | Cost scales with prompt and output length |
| Batch inference | Per token | Set per model on its batch listing | Asynchronous jobs for work that can wait | Results are not immediate |
| On-demand GPU (RTX 4090, H100, H200) | Per hour | Set per GPU type | A dedicated instance you control | Billing continues while the instance is on, busy or not |
| Spot GPU | Per hour | Set per GPU type, below on-demand | Same hardware class at a reduced rate | Capacity can be reclaimed |
| Serverless GPU | Per second | Set per GPU type | Billing only while a request runs | Cold starts and per-request overhead |
Only the hourly rows bill you for time you aren't using.
LLM API prices per million input and output tokens
Of Novita's roughly 200-model catalog, BenchLM counts 93 as text LLMs, and 60 of those are paid per token.
So "200 models" overstates what a chat-completion buyer can choose from by more than three to one.
Swap a long-prompt, short-answer job for the reverse and the cost changes at the same model, because Novita prices input and output separately.
When a retry loop in a pipeline fires again on an image, video or audio model, spend multiplies directly. Those models are billed per unit generated, not per token, so your bill follows the number of outputs.
If you are running this arithmetic for your own team, see what the same workload costs on Activepieces.
Novita AI free credits, rate limits, and discounts
Unless your workload runs on the meter a discount covers, that discount does nothing for you, because Novita AI's discounts attach to specific meters rather than to the whole bill.
Novita AI free credits for new accounts
If you are opening a new account, you can get starter credit, and referral and startup-program credit exist as separate routes. Credit draws down at the same per-unit rates as paid usage, so it buys a fixed quantity of tokens, GPU hours or serverless seconds.
Use the credit to measure, not to build on. Run your real workload long enough to see tokens per request or seconds per job, then take those figures into the utilization arithmetic. Credit that expires unused measured nothing.
Check the current amount and expiry on Novita AI's own pricing and docs pages before you plan around it, because promotional terms change.
Rate limits by usage tier
Rate limits apply to the token meter, the one where you pay per request. They're set per account and per model, and they loosen as your usage and spend history grow.
A new account can therefore be throttled well below what its balance would pay for. A batch job or a launch-day spike can hit the ceiling before it hits the budget.
If your workload is bursty, the limit is your real capacity, not the price list.
When a ceiling is too low, two exits exist: ask for a higher tier, or move that traffic to a GPU or serverless meter, where you rent capacity instead of sharing a request quota.
Novita AI batch, spot, and cache discounts
Three kinds of discount exist, each tied to one meter:
- Batch pricing on the token meter, for work that can wait for results instead of needing an immediate reply.
- Spot GPU pricing on the GPU-hour meter, which is cheaper because the capacity can be reclaimed.
- Reserved or committed pricing on GPU hours, which trades flexibility for a lower rate.
The conditions decide the value. Batch only helps if the job tolerates delay. Spot only helps if the job survives interruption, and a restart costs you the hours already run.
Commit to reserved pricing and the commitment bills whether or not you use it, so it only helps if the instance stays busy.
For each one, read the conditions on the relevant Novita AI docs page and price it against your own utilization, not the headline rate.
Novita AI monthly cost example for 20,000 conversations
Run the same 20,000 support conversations a month through each meter and the amounts come out very different. Tokens scale with use, an always-on H100 bills for every hour regardless, and serverless GPU bills only for seconds it runs.
Token-billed API: the bot, line by line
On the token meter, Llama 3.3 70B on Novita AI is priced at $0.135 per million input tokens and $0.40 per million output tokens.
Output costs roughly three times as much as input, so a bot that writes long answers costs more than one that reads long prompts.
Count both sides separately for each conversation.
If your bot resends the full chat history on every turn, input tokens grow with each message, and that's the line that surprises people. Multiply the per-conversation total by 20,000 and the bill scales up with traffic.
Put the same model behind a trigger and the token line is capped by how you build the workflow. An Activepieces flow that fires on each incoming ticket and calls Novita AI's API once to summarise it spends a bounded number of tokens per ticket.
Activepieces' published pricing page lists 1 credit per flow run, whatever the number of steps, so a two-step flow and a ten-step flow cost the same.
Zapier's pricing page bills per task and Make's bills per module, so splitting the work into smaller, more reliable steps costs extra there, while here the run is the meter and the steps inside it are free.
On the Plus plan, $20 a month covers 10,000 credits, then $0.007 a credit over, so the workflow line stays flat while the token meter does the varying.
Dedicated GPU running 24/7 versus serverless
Three meters, one workload, all in the table below. The token and serverless bills are formulas driven by usage, while the H100 is a fixed number that usage can't lower.
| Assumption | Quantity | Rate | USD |
|---|---|---|---|
| Llama 3.3 70B input tokens | 20,000 conversations × input tokens each | $0.135 per 1M | tokens ÷ 1M × $0.135 |
| Llama 3.3 70B output tokens | 20,000 conversations × output tokens each | $0.40 per 1M | tokens ÷ 1M × $0.40 |
| Serverless GPU | seconds actually executing | per-second rate on the docs page | seconds × rate |
Even if the bot answers nothing at 3 a.m. or all weekend, the H100 line is $2,474.70. Idle hours are billed exactly like busy ones.
Novita AI tokens vs GPU break-even point
Unless your token bill would exceed $2,474.70 a month, the always-on GPU loses. Below that, you're paying for capacity you don't use, and the token meter is cheaper.
To find your own crossover, fill in the token rows from a week of real logs and compare the result to that fixed figure.
In between sits serverless. It removes the idle charge, but you pay per second of execution, so slow cold starts and long generations show up on the bill. Price it from measured seconds per conversation, not from the headline rate.
Worth checking against a plan that does not meter every step: one credit covers a whole run on Activepieces.
Novita AI hidden costs and limits
Price the meter you're running and you've priced only the headline rate. The charges that surprise people sit around it: idle GPU time, storage, data movement, cold starts, balance rules, model retirements and throttling.
Read the unknowns as questions to send to support, not as reasons to assume zero cost.
Novita AI idle GPU and storage charges
When a machine is left on overnight, it bills like one under load, because a GPU-hours meter charges for time an instance is running.
Whether stopped instances keep billing for attached storage or volumes, and what the terms are for egress and image pulls, I couldn't confirm. Treat all three as unknown and price them as a risk, not as free.
Novita AI serverless cold starts and throttling
Because serverless meters by the second, cold-start time is the thing to measure.
If a worker must load a model before it answers, that load time may or may not be billed, and the effect on your per-request cost depends on how often workers scale to zero.
I found no confirmed statement on either point, so test it with your own traffic.
Rate limits are the same story: they exist on any shared inference API. I couldn't verify the thresholds that apply to your account tier, so a burst workload should be load-tested before launch.
Novita AI prepaid balance and model changes
You fund an account and draw it down; that is what pay-as-you-go implies. What I couldn't confirm is any minimum top-up, any expiry on unused credit, or what happens to running jobs when the balance reaches zero.
Ask about all three, because a drained balance that stops a production endpoint is an outage, not a billing detail.
When hosted catalogs of open models turn over as new versions arrive, a model you built a prompt around can be retired. I didn't find a documented notice period, so pin your dependency on one model name and keep a tested fallback.
Novita AI vs Together, Fireworks and RunPod
At $1.45, Novita AI has the lowest H100 hourly rate of the four, the other three all cost more per hour, making Novita AI the obvious choice for cost-conscious projects. Whether that wins depends on which meter your workload runs on.
It also depends on how many hours the GPU sits idle.
Before comparing rates, read the table, which shows which of the three meters each vendor leads with. A per-token price and a per-hour price can't be set side by side until you know your traffic.
| Vendor | Meter you pay on | Who it suits |
|---|---|---|
| Novita AI | Tokens, GPU hours, serverless seconds | Teams that want to start on tokens and move to rented GPUs later |
| Together AI | Tokens | Teams that only call hosted models |
| Fireworks AI | Tokens | Teams that only call hosted models |
| RunPod | GPU hours, serverless seconds | Teams that already run their own containers |
Only Novita AI covers all three meters, so moving between them doesn't mean changing vendor.
Per-token APIs compared
All three sell hosted open models by the million tokens. Compare the same model, same context length and same input-to-output ratio on each pricing page, since output tokens usually cost more than input.
A chatbot with long replies and a classifier with one-word replies can rank these vendors in opposite orders.
Check rate limits and free credits on each page before you commit; they change often enough that a figure copied into a spreadsheet goes stale quickly.
GPU rental: Novita AI versus RunPod
On the hourly H100, Turion's June 2026 comparison puts Novita AI at $3.39 and RunPod at $3.99, meaning RunPod's premium is $0.60 per hour.
Lambda is $4.59, CoreWeave $6.16 and AWS $6.88, with AWS running roughly 50% above Lambda's cost, meaning AWS-dependent organizations pay substantially more than Lambda deployments.
Read Novita's own bare-metal page as what it is, a vendor's page; it lists the same figures for the four rivals.
$438 a month is the $0.60 gap to RunPod for one GPU left on for 730 hours, showing how hourly differences accumulate over continuous operation, so teams deploying continuous GPU infrastructure should scrutinize hourly rates.
Against AWS the gap is $3.49 an hour, about $2,548 a month.
An idle GPU still bills at full rate, so a cheap hour you leave running costs more than an expensive hour you switch off.
Which one fits which workload
Choose Together AI or Fireworks AI if you only call models and never touch a GPU. Choose Novita AI if your traffic is likely to move between meters, or if steady GPU use makes the lowest hourly rate matter.
An idle GPU still bills at full rate, so a cheap hour you leave running costs more than an expensive hour you switch off.
Choose RunPod if you bring your own container and want per-second serverless billing for bursts.
Whichever you pick, set a shutdown schedule first, because idle hours are what separate these vendors' bills most.
What Activepieces does about this
The problem in this article is that the bill follows the meter, and the meter is set by how often something calls the model and how long capacity sits on. Activepieces works on the first half of that.
It is the layer where you decide when a call to Novita AI happens, so you can bound the token meter by design instead of by hope.
The core is MIT-licensed and ships 747 integrations. If you can't find Novita AI among them, you can generate an integration with AI and add it privately to your own instance.
A flow that fires on each ticket, calls the API once and writes the result back spends a bounded number of tokens per run. Put a filter or a schedule in front of the call and a retry loop can't quietly multiply it.
Activepieces' own pricing is flat per run, which keeps the workflow line predictable beside a token line that isn't.
The pricing page lists 1 credit per flow run regardless of steps, and the Plus plan is $20 a month for 10,000 credits, then $0.007 a credit over. Take the 20,000-conversation example and assume one run per conversation.
That is $20 plus 10,000 over-allowance credits at $0.007, so $90 a month for orchestration, however many steps each flow has. Token spend comes on top of that and still varies with prompt and answer length.
It does not fix the idle-GPU meter, and I would rather say so than imply it does. A flow can't make a dedicated H100 cheaper while it is on.
I haven't confirmed that Novita AI exposes an API to stop instances, so check its docs before you build a shutdown schedule around one. Where it does, a scheduled flow is a sensible place to run that schedule.
Until then, treat Activepieces as the way to keep the token meter from running away, and keep the stop-the-GPU habit on your own checklist.
Frequently asked questions
Does Novita AI charge a monthly subscription?
No monthly fee is needed to call its API; you pay for what you meter. That means tokens for model calls, GPU hours for dedicated instances and seconds for serverless.
The bill moves with usage, so a quiet month costs little and a month with a forgotten GPU instance doesn't.
How do I stop a Novita AI bill running away?
Start with a prepaid balance, because it sets a ceiling on what the account can spend. Then stop GPU instances when you finish with them, since an idle instance keeps billing, and use spot capacity only for work that can be interrupted and restarted.
Beyond the balance, you want per-key spending limits and a usage dashboard. I haven't confirmed either from Novita AI's documentation, so check the console before you rely on them.
If per-key limits are missing, use a separate key per project anyway. You can then revoke one key without taking everything down.
Is Novita AI's API compatible with the OpenAI SDK?
Yes, for chat-style model calls. Novita AI exposes an OpenAI-compatible endpoint, so in most cases you change the base URL and API key and leave the rest of your client code alone.
The model names are Novita AI's own catalogue, not OpenAI's, so update them too.
Test any feature beyond plain chat completions before you migrate, because compatibility rarely covers every parameter.
Can I integrate unavailable providers?
Yes. Activepieces runs on an MIT-licensed core with 747 integrations, and where none exists for a provider, you can generate one with AI and add it privately to your own instance, so Novita AI doesn't have to block a project.
I haven't confirmed that its built-in OpenAI step accepts a custom base URL, so if it doesn't, generate a Novita AI integration and add it privately instead.
Trigger a flow often and your spend multiplies, because each workflow run that calls a model is metered in tokens. Put the call behind a filter or a schedule you control, and use the prepaid balance as your backstop.
Related reading
Running the numbers
See what the same workload costs here.
Free forever plan, and every paid plan self-hosts at no extra cost.
See pricing Talk to sales
