> ## Documentation Index
> Fetch the complete documentation index at: https://docs.rockapi.ai/llms.txt
> Use this file to discover all available pages before exploring further.

# Mercury 3 API: integration, pricing checks, and latency testing

> Evaluate mercury-3 on Rock API: verify model access and billing, send a Chat Completions request, and measure latency and task quality before migration.

To evaluate Mercury 3 through Rock API, first look up `mercury-3` in the [current model catalog](https://rockapi.ai/pricing). Confirm that your account and selected group can use it, record the supported endpoint and request limits, then test it against your own workload. A model name in an article does not establish current availability or access for every group.

This guide focuses on integration and evaluation. Use [Inception's official site](https://www.inceptionlabs.ai/) to find the vendor's current model documentation and announcements. Context limits, benchmark scores, release dates, and migration deadlines need a specific current source before you use them in a production plan.

## Check the route before writing code

Record these details alongside your evaluation results:

| Item | What to verify |
| - | - |
| Model | Exact ID `mercury-3`, access for your group, and the model version if disclosed. |
| Endpoint | Chat Completions support and any required or unsupported parameters. |
| Features | Streaming, tool calls, structured output, and cache behavior required by your application. |
| Limits | Input context, maximum output, concurrency, and rate limits for the actual route. |

OpenAI compatibility describes a request format. It does not guarantee that every model implements every OpenAI endpoint or parameter. Read the [model integration guide](https://docs.rockapi.ai/models-and-providers.md) before carrying over provider-specific options.

## Send a minimal request

Follow [authentication](https://docs.rockapi.ai/authentication.md) to configure `ROCKAPI_API_KEY` in your environment. Run this example only after confirming the model and endpoint above; it sends a billable request.

```bash theme={null}
curl --fail-with-body https://rockapi.ai/v1/chat/completions \
  -H "Authorization: Bearer $ROCKAPI_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "mercury-3",
    "messages": [{"role": "user", "content": "Write three checks for an API release."}]
  }'
```

An OpenAI Python client can express the same basic request:

```python theme={null}
import os
from openai import OpenAI

client = OpenAI(
    api_key=os.environ["ROCKAPI_API_KEY"],
    base_url="https://rockapi.ai/v1",
)
response = client.chat.completions.create(
    model="mercury-3",
    messages=[{"role": "user", "content": "Write three checks for an API release."}],
)
print(response.choices[0].message.content)
print(response.usage)
```

Keep the request time, model, group, and request ID when available. If a request fails, consult [errors and rate limits](https://docs.rockapi.ai/errors-and-rate-limits.md) before retrying repeatedly.

## Reconcile pricing instead of assuming a discount

Open the model's details and record currency, price unit, billing mode, input/output categories, and any cache or context tiers. Apply the actual group multiplier only if it is not already included in the displayed effective rate. Rock API does not promise a universal discount across models or requests.

Top-up payment determines the balance credited to your account; it is separate from API consumption pricing. A temporary reservation can also differ from settlement. Follow [pricing and billing](https://docs.rockapi.ai/pricing-and-billing.md) and reconcile a completed request using its saved rates, usage categories, and group. Today's catalog is not a substitute for a historical settlement record.

## Measure latency and useful output

Use the same representative prompts, output requirements, concurrency, and client region for Mercury 3 and your baseline. Separate cold-cache and warm-cache runs when relevant. Record failures and retries rather than discarding them.

For streaming routes, measure time to first content separately from time to complete output. Tokens per second measures only a defined generation interval; it does not include every network, queue, prompt-processing, or tool delay. For agents, measure the entire task and count tool errors and repeated steps.

Compare median and p95 completion times, task success rate, and settled cost per accepted result. Verify factual correctness, output schema, language quality, and tool arguments. Expand traffic gradually only when the candidate meets your own acceptance criteria, with a tested rollback path to the previous route.

## Start an evaluation

[Create a Rock API account](https://rockapi.ai/sign-up?utm_source=docs\&utm_medium=referral\&utm_campaign=mercury_3_api_en), check the model details, and use the [quickstart](https://docs.rockapi.ai/quickstart.md) to prepare a small evaluation. Set a budget and quality threshold before deciding whether to migrate.


This documentation is built and hosted on [Mintlify](https://mintlify.com), a developer documentation platform.