Pricing
Price your agent. Pay for what it uses.
Start from one of the example agents and move the sliders to your volume to see what it costs on Tarka, or with any part in your own infrastructure. The unit prices below come from the billing catalog.
Price an agent like these.
Pick an example to fill in the table, move any slider to your own volume, and choose where each component runs: on Tarka, or in your own infrastructure with a Tarka cell.
| Component | Where it runs | On Tarka | Your infrastructure |
|---|---|---|---|
Agent hostsagent hostOne host for each person | On Tarka€91.2520 × 1 vCPU · 2 GiB · Europe | Your infrastructure≈ €1555 vCPU · 9 GiB · 40 GiB disk | |
Language modelmodel inferenceEvery message and reply | On Tarka€10.9360M in · 4M out · Europe | Your infrastructure€9,198–17,5814× B200 · 614 GB | |
SpeechspeechVoice notes in, spoken replies out | On Tarka€0.003,000 minutes of audio · Free for a limited time | Your infrastructure€263–9861 GPU · 24–48 GB | |
DocumentsdocumentsReading scans and PDFs | On Tarka—not used | Your infrastructure—not needed | |
Code sandboxescode sandboxRunning code the agent writes | On Tarka—not used | Your infrastructure—not needed | |
Batch jobsbatch jobScheduled and batch work | On Tarka—not used | Your infrastructure—not needed | |
Files and sourceobject storagehosted gitBuckets and repositories | On Tarka—not used | Your infrastructure—not needed | |
Cell servicesgatewayGateway and ingress for the parts you run | With any part in yours | On TarkaIncludedpart of the platform | Your infrastructure—only with a part in your infrastructure |
Tarka cellLicense, updates and support | With any part in yours | On TarkaIncludedpart of the platform | Your infrastructure—only with a part in your infrastructure |
| Per monthIn EUR, from the personal assistant example | On Tarka€102.18billed by use | Your infrastructure—nothing runs in your infrastructure |
Agent hosts run in Europe and are billed for allocated CPU and RAM while they run: about €4.56 a month for each of the 20 hosts, before model calls. The model’s tokens cost €10.93 a month on Tarka at this volume, while one copy of DeepSeek V4.1 Flash on your own GPUs rents for €9,198–17,581 a month before staff, monitoring and a second copy. Run the model yourself when rules require it or your volume would keep the GPUs busy; otherwise call it on Tarka.
Tarka amounts use price list 2026-11-04.1: Agent Hosts, batch jobs and sandboxes in Europe by allocated CPU and RAM over running time, models per token, audio second or page; storage is metered but not charged yet. Each example sets the sliders to a rough month of its use and brings its host size, model and tokens per request. Your-infrastructure figures are estimates: one copy of the model on B200 GPUs at $3.50–6.69 an hour and smaller GPUs at $0.40–1.50, other servers and disks at on-demand cloud list prices, with no staff, monitoring or second copy. Tarka’s cell fee is a placeholder. Amounts in currencies other than USD and EUR use market rates of 23 October 2026.
Unit prices.
One account for model APIs, speech, documents and managed compute, billed from prepaid credit. Current rates come directly from the billing catalog.
Catalog 2026-11-04.1 · release verified 2026-11-04 · Live catalog
Prepaid usage, with $5.00 starter credit once per approved account. No subscription is required. EUR amounts use the catalog’s conversion factor; your account’s billing currency applies.
Model pricing
Utilities are free for a limited time. Regular prices are crossed out below; the current price is zero. Inference models and compute products keep their listed prices.
| Model | Input / 1M tokens | Cached input / 1M | Output / 1M | Audio / second | Successful page |
|---|---|---|---|---|---|
deepseek-v4.1-flash | $0.33 | $0.0066 | $1.32 | — | — |
deepseek-v4.1-flash-eu | $0.33 | $0.0066 | $1.32 | — | — |
glm-5.3 | $1.375 | $0.264 | $4.40 | — | — |
glm-5.3-flash | $0.1485 | $0.0297 | $0.495 | — | — |
glm-5.3-flash-eu | $0.1485 | $0.0297 | $0.495 | — | — |
glm-ocr Free for a limited time | — | — | — | — | |
glm-ocr-nepali Free for a limited time | — | — | — | — | |
himalaya-bf16 | Free | — | — | — | — |
himalaya-bf16-private No Tarka content capture | $0.05 | $0.01 | $0.10 | — | — |
himalaya-gemma-4-bf16 | Free | — | — | — | — |
himalaya-gemma-4-bf16-private No Tarka content capture | $0.08 | $0.02 | $0.20 | — | — |
himalaya-gemma-4-q4 | Free | — | — | — | — |
himalaya-gemma-4-q4-private No Tarka content capture | $0.06 | $0.02 | $0.16 | — | — |
himalaya-gemma-4-q8 | Free | — | — | — | — |
himalaya-gemma-4-q8-private No Tarka content capture | $0.07 | $0.02 | $0.18 | — | — |
himalaya-q4 | Free | — | — | — | — |
himalaya-q4-private No Tarka content capture | $0.03 | $0.01 | $0.06 | — | — |
himalaya-q8 | Free | — | — | — | — |
himalaya-q8-private No Tarka content capture | $0.04 | $0.01 | $0.08 | — | — |
kimi-k3 | $2.97 | $0.297 | $14.85 | — | — |
kokoro Free for a limited time | — | — | — | — | |
nepali-parler-tts Free for a limited time | — | — | — | — | |
paddleocr-vl-1.6 Free for a limited time | — | — | — | — | |
piper Free for a limited time | — | — | — | — | |
pp-doclayout-v3 Free for a limited time | — | — | — | — | |
qwen3-asr Free for a limited time | — | — | — | — | |
qwen3-tts Free for a limited time | — | — | — | — | |
qwen3.8-flash-next | Free | — | — | — | — |
whisper-large-v3 Free for a limited time | — | — | — | — | |
whisper-nepali-medium Free for a limited time | — | — | — | — |
Speech recognition is measured by processed input audio; speech generation by output audio duration, including generated audio whose delivery began. Documents are measured by successful pages. Utility calls do not add token charges. Private endpoints retain billing metadata.
EU compute pricing
| Product | vCPU / 720 hours | GiB RAM / 720 hours | 1 vCPU + 1 GiB / 720 hours |
|---|---|---|---|
| agents | $3.00 | $1.00 | $4.00 |
| jobs | $4.50 | $1.50 | $6.00 |
| sandboxes | $6.00 | $2.00 | $8.00 |
720 hours is a comparison period, not a monthly fee or cap. Charges use allocated CPU and RAM over running time. Existing grandfathered Agent Hosts remain free. New hosts stop when prepaid credit runs out and can be resumed after a top-up; their data is retained.
Compute · US East
US East has separate prices. The total depends on your size and its minimum hourly charge. Your existing EU resources keep their current prices.
| Product | 1 vCPU + 1 GiB RAM + 1 GiB scratch / hour | Minimum running time |
|---|---|---|
| agents | Quote your stack | 60 seconds |
| jobs | $0.1511443 | 60 seconds |
| sandboxes | $0.1511443 | 60 seconds |
The hourly examples include the minimum charge for that size. Agent Hosts are quoted per stack, including all application and platform workloads. Larger CPU, memory or scratch allocations need a new quote. No runtime charge applies if execution never starts. Network access and persistent storage are not included in these examples; their prices and availability must be confirmed before creation.
Estimate your usage
Actual usage is rounded per request or cumulative compute runtime. This estimate excludes free starter credit.
Create an account and price your first agent.
Sign up, then request the products you need from the console.