AI inference provider

OpenAI-compatible AI inference API

Two frontier models on dedicated deployments, billed per token and spoken over the OpenAI Chat Completions API. Point an existing OpenAI client at the base URL below and change nothing else.

We are a proxy in front of dedicated model deployments we operate and pay for. We train nothing, we store no prompts, and every rate on this page is read live from the same table the invoice is computed from.

Models

Two models, both on the same wire

Both are reasoning models served from dedicated capacity in US East 2. Text is the only input modality on either.

Opus speaks a messages-style API upstream rather than Chat Completions. We translate to it in front of the model and back again, so a client only ever writes OpenAI Chat Completions. Which platform each deployment runs on is not published here; what it accepts is, and that is the part your code has to agree with.

Pricing

Per token, published without a key

Prices are per million tokens in US dollars. The table below is rendered from /api/openrouter/v1/models — the same rate document the invoice is computed from — so it cannot drift from what you are billed.

Static values above were verified against the live rate document on 20 August 2026.

Reasoning tokens are output tokens and are billed once at the output rate, not twice and not extra. A discounted cached-input rate exists for eagleai/gpt-5.6-sol and only for it: that upstream caches prompts on its route automatically, while nothing on the Opus route ever sends a cache flag, so a cached rate there could never be earned.

API

Change the base URL, keep your client

POST /api/openrouter/v1/chat/completions, OpenAI Chat Completions in and out. Authenticate with Authorization: Bearer <key> or X-API-Key: <key> — the same issued key, either header.

curl https://eagleai.online/api/openrouter/v1/chat/completions \
  -H "Authorization: Bearer $EAGLEAI_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "eagleai/gpt-5.6-sol",
    "messages": [{"role": "user", "content": "one sentence on tides"}],
    "stream": true
  }'

from openai import OpenAI

client = OpenAI(
    base_url="https://eagleai.online/api/openrouter/v1",
    api_key=os.environ["EAGLEAI_KEY"],
)
client.chat.completions.create(
    model="eagleai/claude-opus-5",
    messages=[{"role": "user", "content": "one sentence on tides"}],
)

What each model actually accepts

Every row here was measured against the live deployments rather than copied from an upstream's published spec. A parameter this surface cannot honour is refused by name with a 400 and error.param set, never dropped quietly.

Both models are reasoning models, and reasoning models fix their own sampling. Sol takes reasoning_effort and forwards it unaltered; minimal is the one value its deployment refuses. The reasoning object is rejected as an unknown field on Sol, so neither model is advertised as taking it.

On Opus, reasoning runs by default and its tokens are billed as output tokens. You can switch it off — reasoning: false, {"enabled": false} and {"exclude": true} all work — but you cannot turn it up: asking for an effort level or an explicit thinking budget is a 400. Measured on 20 August 2026, the same prompt answered with reasoning left on cost 197, 182 and 184 output tokens across three runs against 137, 141 and 141 with it off, so the switch is real and it moves your bill.

Capacity

The quota is the deployment's, and we publish it

These are each deployment's own provisioned limits rather than a number we chose, and all three are enforced here before the upstream is contacted. Over any of them is an immediate 429 carrying Retry-After. Nothing queues.

The throughput and latency columns are our own measurements, taken on 20 August 2026 against the live surface with cold, uncached prompts and zero errors: Sol at 16 concurrent over 180 seconds, Opus at 2 concurrent. They are what we recorded on that run, not a guaranteed rate — we do not offer a service level agreement, and the terms say so plainly.

The token budget is a ceiling on the deployment, not an allowance handed to each key: every caller of one model draws on the same minute. A request is refused once that minute is already spent, and never for being large.

Access

Getting a key

Data handling

What a request leaves behind

Prompts, messages and completions are held in memory only for as long as it takes to forward them, and are never written to disk, to a database, or to our logs. We operate no models of our own and fine-tune nothing. One metering record per request — identifiers, token counts, timings and money, no content — is kept for 365 days and then deleted by the server.

Requests are served from dedicated deployments we operate on a third-party cloud AI platform, whose abuse monitoring may retain prompts and completions for up to 30 days. Modified abuse monitoring (zero data retention) is not enabled on these deployments, and we publish "zdr": false in the catalogue rather than claim a posture we do not hold. If zero retention is a requirement for you, this API is not currently a fit. The full account is in the privacy and data retention notice.

Also from EagleAI

Computer vision platform

Separately from this API, EagleAI builds a computer-vision platform that turns existing cameras, NVRs and operational feeds into real-time perception and action infrastructure. It is a different product with a different deployment model, and nothing on this page applies to it.