LLM API quickstart

Choose a model for your application

Compare upstream capabilities in the model library. Catalog identifiers are proposed request identifiers; use the exact ID returned by your enabled API model list. Upstream context windows do not establish the deployed context or maximum output length.

• MiMo-V2.5 — xiaomi/mimo-v2.5
• Hy3 — tencent/hy3
• MiniMax-M3 — minimax/minimax-m3
• GLM-5.3-Flash — z-ai/glm-5.3-flash

Configure your server environment

Set PROBSEEK_CHAT_ENDPOINT to the full HTTPS chat-completions URL supplied for your account, including its path (typically /v1/chat/completions), and PROBSEEK_API_KEY to your API key. This is a full request URL, not an SDK base URL. Keep keys out of browser code, source control, URLs, and logs. Install the Python dependency with python -m pip install requests, or use cURL in a shell.

Evaluate responses before production use

Start with non-sensitive test input. Inspect choices, finish_reason, and usage, and handle HTTP errors and timeouts. Confirm prices, enabled modalities, tool support, context and output limits, and effective Privacy & Data Policy before sending production data.

Python and cURL request examples

Illustrative request only. Confirm the full endpoint, enabled model ID, supported fields, and response format before running it. Copying or replaying an example on this website does not call an API.

import os
import requests

response = requests.post(
    os.environ["PROBSEEK_CHAT_ENDPOINT"],
    headers={"Authorization":
        "Bearer " + os.environ["PROBSEEK_API_KEY"]},
    json={
        "model": "xiaomi/mimo-v2.5",
        "stream": False,
        "max_tokens": 256,
        "messages": [{"role": "user",
            "content": "What could I build with AI?"}]
    },
    timeout=60
)
response.raise_for_status()
print(response.json())

curl "${PROBSEEK_CHAT_ENDPOINT:?Set your full HTTPS chat endpoint}" \
  --fail-with-body --max-time 60 \
  -H "Authorization: Bearer $PROBSEEK_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "xiaomi/mimo-v2.5",
    "stream": false,
    "max_tokens": 256,
    "messages": [{
      "role": "user",
      "content": "What could I build with AI?"
    }]
  }'

For model-specific examples, open a model detail page. For access and pricing questions, read the ProbSeek FAQ.

Access and compatibility status

This website does not publish a verified production endpoint, live model inventory, or compatibility results. A configured console link or example URL alone does not establish service readiness. Self-service access depends on the account service linked by the site; without that link, this website has no registration flow.

Confirm the contract for POST /v1/chat/completions and GET /v1/models with the service operator. The /models/ page on this website is an HTML catalog, not the API model-list endpoint. OpenRouter provider acceptance and listing have not been established by this website.

OpenRouter provider integration requirements

OpenRouter requires streaming chat completions and token usage in both response modes. ProbSeek compatibility and OpenRouter listing are unverified here.

New provider inventories use schema_version 2.4: exact callable IDs, typed input_modalities and output_modalities, deployed limits, and supported_parameters. Token prices belong under their modality as USD cost_usd strings per token, not per million tokens. Root prices are request-scoped. The legacy flat format is for existing integrations. Validate the live inventory against the linked schema; this website does not provide one.

Declare physical datacenters and verified compliance separately. A deployment_region label alone does not establish geographic routing. Keep unlaunched endpoints is_ready: false; omitting it permits automatic launch after OpenRouter checks. No ZDR or certification claim is established here.

The application requires monthly invoicing; the provider guide also allows auto top up. Confirm the payment arrangement with OpenRouter. This website does not establish either payment method.

Non-streaming responses and token usage

For the illustrated contract, send model and messages as JSON with Bearer authentication. With stream: false, inspect choices[].message and choices[].finish_reason. Tool calls may appear instead of text. A length finish reason indicates that the generation limit was reached, not necessarily a complete answer.

Check usage.prompt_tokens, usage.completion_tokens, and usage.total_tokens. Do not estimate billable usage from characters. Ask how cached input, reasoning tokens, failed requests, and interrupted generations are counted; a missing usage object is not evidence of zero cost.

Streaming and completion accounting

Where the endpoint confirms support, stream: true requests Server-Sent Events (SSE). The example below also requests stream_options.include_usage. Parse data events, ignore SSE comments, and handle the [DONE] sentinel separately from JSON. Deltas may contain text or tool-call fragments.

Collect the final usage event before closing the stream; it may have an empty choices array. Treat an error event or a disconnected stream as incomplete even after HTTP 200. Do not assume that cancellation stops billing. Confirm the provider’s stream contract and accounting behavior with test traffic.

Do not stop reading at the first finish_reason: usage can arrive afterward. OpenRouter currently includes a content-free choice and repeats the terminal finish_reason in its chat-completions usage frame, whereas an OpenAI-style usage frame can have choices: []. Accept either shape and account for usage once. Buffer complete SSE events, including multi-line data fields, before parsing JSON; network chunks are not event boundaries.

Illustrative request only. Confirm the full endpoint, enabled model ID, supported fields, and response format before running it. Copying or replaying an example on this website does not call an API.

curl "${PROBSEEK_CHAT_ENDPOINT:?Set your full HTTPS chat endpoint}" \
  --no-buffer --fail-with-body --max-time 60 \
  -H "Authorization: Bearer $PROBSEEK_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "xiaomi/mimo-v2.5",
    "stream": true,
    "stream_options": {"include_usage": true},
    "max_tokens": 256,
    "messages": [{
      "role": "user",
      "content": "What could I build with AI?"
    }]
  }'

Errors, timeouts, and retries

For an OpenAI-compatible endpoint, inspect both HTTP status and the error payload. Correct invalid input (400), credentials (401), payment (402), geographic restrictions (403), missing models (404), or oversized payloads (413) before retrying. Endpoint-specific behavior requires confirmation.

For rate limits (429) or transient server failures (5xx), use bounded backoff with jitter and honor Retry-After when supplied. Set connection and read deadlines. Replaying a request after a timeout or partial stream may create another billable generation; do not automatically replay tool side effects. Use request IDs and non-content metadata for diagnostics, keeping keys and prompt content out of routine application logs.

Model limits, pricing, and data terms

For each deployed model, obtain supported input/output modalities and parameters, context and output limits, serving precision, physical datacenter locations, and rate limits. A model author’s advertised capabilities or recommended configuration does not confirm ProbSeek support. No residency guarantee covering all data, uptime SLA, or measured performance is published here.

Inference currently runs in the United States (US). European Union (EU) inference is planned and is not currently available.

Input and output prices in USD, cache or other charges, and payment terms are not published here. OpenRouter provider billing requires an agreed automated payment arrangement; this website does not establish invoicing support. Review the Privacy & Data Policy for prompt/output logging, retention, training use, recipients, and deletion terms. For retention and recipient details specific to your deployment, contact us before submitting data with particular handling requirements.

ProbSeek does not use API inputs or outputs to train models.

ProbSeek does not log prompt or response content in routine inference logs.

Operational metadata logs are retained for 7 days after creation and then deleted. They contain request IDs, model identifiers, token usage, latency, and status codes, not prompt or response content.

Read the Privacy & Data Policy

Reliability and support

OpenRouter measures successful-request uptime, time to first token, and throughput. These routing metrics are not a ProbSeek SLA. Live measurements, overload and cancellation behavior, and escalation arrangements still need verification.

For access, integration, or billing questions, contact [email protected]. Include a request ID, model ID, UTC timestamp, and status code when available. Do not send API keys or prompt and response bodies in routine support reports.