Java AI API client (India)

Not every Java backend uses Spring — plenty of batch jobs, Android-adjacent backends and plain Maven services call AI models through a bare Java OpenAI SDK instead of a framework abstraction. Those SDKs already expose a configurable base URL, which is all a Java service needs to move from an internationally billed host to unoblox's rupee-billed, OpenAI-compatible endpoint.

Configuring the client

OpenAIClient client = OpenAIOkHttpClient.builder()
    .baseUrl("https://api.unoblox.ai/v1")
    .apiKey("ub-gw-xxxxxxxxxxxxxxxxxxxxxxxxxxxx")
    .build();

ChatCompletionCreateParams params = ChatCompletionCreateParams.builder()
    .model("openai/gpt-5-mini")
    .addUserMessage("Classify this ticket as billing, technical, or other.")
    .build();

ChatCompletion completion = client.chat().completions().create(params);
System.out.println(completion.choices().get(0).message().content());

The builder pattern, the response types and the blocking or async call shape all stay as documented — baseUrl and apiKey are the only two lines that change.

Choosing a model for a backend service

Model id	₹ / 1M input	₹ / 1M output
openai/gpt-5-mini	₹25.20	₹201.60
qwen/qwen3-235b-a22b-instruct-2507	₹9.07	₹55.44
moonshotai/kimi-k2.7-code	₹68.50	₹342.70
qwen/qwen3-1.7b (free)	₹0.00	₹0.00

moonshotai/kimi-k2.7-code is a reasonable default to evaluate for code-heavy Java services given its naming, but validate it against your own workload before committing — don't assume a model's fit from its name alone. Anything not listed here should be priced from its live /models page.

Thread pools and batching

Java backends often fan a single incoming request out into several model calls (classification, then extraction, then a summary). Because unoblox meters by token rather than by request, batching those calls into fewer, larger prompts where your logic allows it reduces overhead without changing the per-token rate you pay. If you're running a fixed-size thread pool for outbound HTTP calls, the same connection-pooling settings you'd use against any HTTPS API apply unchanged here.

Dependency management

Add whichever Java OpenAI-compatible SDK your team already uses (via Maven or Gradle) — unoblox doesn't require a fork or a vendor-specific artifact, since the wire format is the standard OpenAI chat-completions shape.

The invoice difference

Nothing about the Java code implies a currency. What changes is that unoblox settles in rupees on a monthly GST invoice from an Indian entity, claimable as input tax credit, with no international card needed to get a key issued.

Frequently asked questions

Which Java SDK should I use? Any Java client built for the OpenAI chat-completions API works, since that's the protocol unoblox implements — set baseUrl and apiKey to unoblox's values.

Does tool/function calling work? Yes, unoblox implements the same tool-calling contract, so SDK-level function-calling builders work unchanged.

Can I run this from a Spring Boot app too? Yes, though if you're already on Spring AI, its dedicated starter (see our Spring AI guide) is a slightly cleaner fit.

Is there a free model for local development? Yes — qwen/qwen3-1.7b is priced at ₹0.

How do I keep the key out of source control? Load it from environment configuration or a secrets manager at startup, the same way you'd handle any other API credential.

How is usage billed? In rupees, on a monthly GST invoice from an Indian entity, with input tax credit claimable.

Get started in rupees → https://unoblox.ai/sign-in