Scale-to-zero platforms for autonomous agents
Compare scale-to-zero platforms for autonomous agents across European cloud providers, and calculate the cost of cold starts versus idle capacity.
In this article7

What scale to zero means in each hosting shape
Scale to zero describes a platform that reduces its bill to nothing when no work is running. Agents idle for most of the day and then burst, so the term carries a different meaning in each of the three hosting shapes an engineering team can pick, and the differences decide the bill.
- Per-token model API: the provider keeps shared model capacity running and bills only for tokens processed. An idle agent incurs no idle bill, because there is no dedicated hardware to keep warm.
- Serverless container platform: the platform spins the agent's own container down entirely when inactive, so no running replica exists and nothing bills while the agent waits. The trade is a provisioning delay on the next run.
- Reserved endpoint: a replica stays warm continuously. The agent never scales to zero, so the hourly rate applies around the clock whether or not a request arrives.
The first shape removes the idle bill, the second removes the running replica, and only the third guarantees neither. A team whose agents sit idle between runs pays Lyceum only for the tokens a run actually consumes, with no reserved capacity to keep warm: its serverless inference endpoint lists GLM-5.3 Flash at $0.20 per 1M input and $0.50 per 1M output tokens, all prices in USD per 1M tokens excluding VAT.
Cold start as the trade you are making
Cold start is the delay between a request arriving and the first token of the response. Each hosting shape handles it structurally, and the structure, not any benchmark figure, is what a team is buying.
- Per-token API: requests land on shared capacity that is already serving other tenants, so no user-facing cold start arises. The cost is tenancy: performance varies with platform load, and T-Systems notes for its shared LLM serving that realised throughput and time-to-first-token vary with load and sit outside the SLA[1].
- Serverless container: waking a scaled-down replica means provisioning compute, restoring the image and reloading model state before the first token. The delay is structural and recurs after every idle period.
- Reserved endpoint: the warm replica answers immediately, which is precisely what the continuous hourly cost buys. T-Systems positions its dedicated serving for strict latency SLAs, bursty workloads and single-tenant infrastructure[1].
Pricing one idle-heavy agent three ways
Consider an agent idle 90 percent of the day. Under a per-token API the idle 90 percent costs nothing and the bill is tokens alone. Under a serverless container the idle time also bills nothing, with wake-up delay as the price. Under a reserved endpoint the full 24 hours bill at the hourly rate. For an agent needing its own container, Lyceum bills serverless GPU execution per second, at $1.69/hr for an L40S and $3.59/hr for an H100. Other providers' published rates start from $0.02 per 1M input tokens.
| Provider | Cheapest listed per-token input rate | Reserved GPU hourly rate | Scale to zero |
|---|---|---|---|
| Lyceum | $0.20 / 1M input (GLM-5.3 Flash) | – | Yes, per-token endpoint |
| IONOS | $0.11 / 1M input (Qwen3.5-9B) | – | – |
| Nebius | $0.06 / 1M input (Nemotron-3-Nano-30B-A3B) | from $1.55/hr (L40S, on-demand) | – |
| Open Telekom Cloud | €0.20 / 1M input (Standard tier) | – | – |
| OVHcloud | – | – | – |
| Scaleway | €0.15 / 1M input (gpt-oss-120b) | €0.93/hr (L4 managed inference) | – |
| STACKIT | €0.15 / 1M input (standard tier) | – | – |
| TensorX | $0.02 / 1M input (Qwen3-Embedding-8B) | – | – |
A "–" means the provider's page lists no value. The rates are read from each provider's own public pricing page on 22 September 2026 and are not like-for-like: model tiers, currencies and billing units differ, and a reserved hourly rate covers a GPU rather than tokens. To price the same day of traffic under each shape, multiply the day's input and output tokens by the per-token rates, and compare that against the full 24 hours at the reserved hourly rate.
When reserved capacity wins anyway
Reserved capacity remains the right choice in three situations. First, when the agent's response must begin immediately after every idle gap, because a scaled-down replica cannot promise that. Second, when traffic is dense enough that the reserved hourly rate falls below what the same tokens would cost per-token, which happens once utilisation rises past the idle-heavy shape this article prices. Third, when the workload needs single-tenant infrastructure or contractual latency guarantees, which T-Systems reserves for its dedicated serving[1]. Nebius prices commitment discounts at up to 35 percent below on-demand rates for multi-month reservations[4], and Scaleway offers dedicated deployments at the hourly rate listed above[6].
Choosing for your own traffic shape
The decision reduces to three measured inputs: the share of the day the agent is idle, the tokens each run consumes, and the wake-up delay the workload tolerates. An idle-heavy agent with a tolerable delay fits a per-token API, where the idle share of the day costs nothing. An agent that needs its own container and can wait out a wake-up fits a serverless container platform billed per second. Re-check the rates against each provider's page at the date of the decision, because every figure in this article carries a retrieval date and none is guaranteed to hold.
Frequently asked questions
What does scale to zero cost?
True scale to zero costs nothing during idle periods. On a per-token API, an idle agent incurs no bill because you pay only for processed tokens. On a serverless container platform, billing stops when the replica spins down, though you pay per second while it processes active bursts.
Do serverless GPUs have cold starts?
Yes, serverless GPU containers incur structural cold-start latency. When an agent wakes up from an idle state, the platform must provision a new replica and load the model into memory, which introduces a delay before the first token is generated.
Is a per-token API cheaper than a reserved endpoint for agents?
For agents that idle for long stretches, a per-token API is significantly cheaper. A reserved endpoint charges a continuous hourly rate regardless of utilisation, whereas a per-token API bills only for the actual prompts and responses the agent processes.
How do I price an idle agent?
To price an idle agent, map its traffic shape. Calculate the cost of the active bursts using per-token rates or per-second container billing, and add the continuous hourly cost of any reserved capacity required to keep the endpoint warm between runs.
Can reserved capacity scale to zero?
No, reserved capacity by definition cannot scale to zero. You pay a continuous hourly rate for the allocated hardware to guarantee it remains warm and available instantly, making it expensive for agents with long idle periods.
Sources
Rankings and scores on Optinest cannot be bought — no commercial relationship changes how a provider is assessed. How we score
Related reading
All Agents →Inference APIs with fast cold starts for bursty traffic
Which inference platforms scale to zero, keep replicas warm or bill per token, what idle capacity costs, and how to choose for bursty LLM traffic.

Private LLM hosting with autoscaling and scale-to-zero
Which platforms host a private LLM with autoscaling and scale-to-zero, how they bill, what idle GPUs cost, and how to choose for bursty or steady traffic.

Deploying a Hugging Face model as an API endpoint
Evaluate the routes to deploy a Hugging Face model as an API endpoint. Compare managed endpoints, serverless platforms, and VMs across European providers.
