Subscriptions for running repeat evaluations on your own test sets, across the evaluation areas available today: AB-VietLLM, AB-Doc, AB-Voice and AB-Memory. The prices are being tried with the first pilot group.
All three paid plans share the evaluation areas available today and differ in scale, automation and governance. You are not charged a separate fee for each area.
See the public benchmarks. The trial workspace is being built.
A small team starting systematic testing.
Repeat testing, teamwork, and catching errors before release.
Many systems, with a need for governance and larger volume.
Priced to the deployment and support you need.
Each area shows a plain usage unit. The way it is counted below is planned and will be fixed together with the rate card after the pilot.
| Code | Area | What it evaluates | Usage unit (planned) |
|---|---|---|---|
| AB-VietLLM | Language models | Compare models on Vietnamese tasks: quality, instruction following, tool use, speed and cost | Test cases, models, actual tokens |
| AB-Doc | Document processing | Accuracy of the important fields, effort to check and fix, processing time and cost | Pages processed, solutions compared, scoring configuration |
| AB-Voice | Voicebots | Task completion, serious errors, latency, interruptions, handover to a person | Simulated minutes and audio minutes evaluated, counted separately |
| AB-Memory | Long-term memory | Long context (measured on the base model) and recall across sessions (measured on the whole system) | Scenarios, sessions and turns, tokens |
Access to the platform and the amount you use are kept apart, so you can see what you are paying for.
By default runs stop when the included budget is spent. You choose to allow overage and set a ceiling.
Before each run the interface shows the number of executions and an estimate. One test case run on 3 models, repeated 5 times, is 15 executions.
Fees charged directly by the model provider are then shown apart from AsiaBench’s fee.
Published after the pilot, once the cost of building inputs, calling the system under test, scoring, rerunning and storing has been measured.
The real rate card is not published, so this only shows how the sum works. Overage only arises when you choose to allow it. Not included: amounts charged directly by third parties (if you use your own API key) and tax where it applies.
A plan includes guides and support for using the product. It does not include unlimited advice, labelling or manual analysis.
Three kinds of result are kept distinct, so a reader knows where each number comes from.
Run by AsiaBench under its published process. Always free.
A report within the contract scope, with a named person responsible for the method.
Shows the configuration and data source. It is not automatically an independent certification and does not carry AsiaBench Verified.
Business, vendor or model lab: leave your email and we’ll get back to you within a few business days.