Skip to contentQloudOpen models
Console login

Inference for asynchronous agents

Intelligence in motion.

Off-peak inference

Patience is 70% cheaper.

Batch runs agents and workflows at 70% lower token rates than Realtime. Need results sooner? Async is 40% lower.

Start building
Qloud scheduler
Realtime demandBatch workIdle GPU capacityRealtime demand rises and falls, leaving idle GPU capacity. After the scheduler, Batch work fills that spare capacity.

Backed by

Pricing / 01

Same job. Three prices.

One GLM-5.3 job: 1M fresh input, 1M output tokens. Pick a tier.

Batch inference

$1.20per completed job

$2.80 saved vs Realtime · 70% lower

For latency-tolerant agents and workflows with room to wait.

Input
$0.27
Output
$0.93
Cached input
$0.078

Per 1M tokens, GLM-5.3 list rates.

Batch: $1.20 per completed job.

Workloads / 02

Work no one is waiting on.

Most agent work runs in the background. Stop paying for speed it never uses.

  • Research agents

    Investigate, compare and assemble findings across many sources in parallel.

    BatchAsync for earlier results
  • Background coding

    Code reviews, maintenance and tool-driven workflows where each step depends on the last.

    AsyncBatch for independent reviews
  • Data pipelines

    Extract, enrich, classify and analyze records, sentiment included, in bulk.

    BatchAsync for background processing
  • Evaluations & generation

    Evaluation runs, simulations and synthetic data that can finish within a longer window.

    Batch

How Batch works / 03

Hand off the work. Collect it finished.

Runs as capacity frees up. Returns with the cost of every job.

Example batch

Illustration: 1,000 GLM-5.3 jobs are submitted on Batch and run in groups as capacity frees up. They finish with $1,200.00 billed, $1.20 per completed job, and $2,800.00 saved against Realtime.

Economics / 04

Why work that can wait costs less.

The discount comes from how the work is served. It does not depend on cache hits.

  • Scheduling freedom

    Run when capacity is free.

    Work with room to wait fills the gaps between realtime peaks, so less capacity sits idle.

  • Fuller batches

    More useful work per pass.

    Flexible admission keeps batches full, so each GPU pass does more completed work.

  • Context reuse

    Skip repeated input work.

    Shared prefixes and warm sessions avoid recomputing context where capacity allows.

Silicon / 05

GPUs today. Our own silicon next.

Qernel is developing inference chips built on charge-domain compute (cDP) to push the cost of each workload lower again. Qloud runs on GPUs today.

Explore Qernel

Open models / 06

Open models. Three tiers.

Tell us the model, the volume and when you need results. We will show you what it costs to complete on each tier.

Explore models and pricing