Agnes AI offers a polished API gateway but provides no proof its models are truly in-house

Agnes AI is an API gateway and model catalog that exposes a family of in-house multimodal models — text, image, and video — through an OpenAI-compatible interface. The pitch is simple: swap the base URL, keep your existing OpenAI client code, and gain access to Agnes's own model lineup without rewriting integration logic.

The current catalog spans five models. On the language side, agnes-2.5-flash and agnes-2.0-flash handle chat, tool calling, streaming, and vision inputs. The 2.5 variant advertises a 512K context window with up to 65.5K output tokens; the 2.0 model was rolled back from a temporary 1M context window in June 2026 and now sits at 256K context with 64K max output. Two image models cover text-to-image and image-to-image workflows with URL or Base64 output. The video model, agnes-video-v2.0, is asynchronous — you POST a generation task, receive a video_id, and poll separately for the result. The README is explicit that task_id is deprecated for result queries; use video_id only.

Rate limits are tiered across Free, Enterprise, and Token Plan tiers. Free users get 20 requests per minute on text models and 1 RPM on video. Token Plan subscribers jump to 1,000 RPM for text. Subscription quotas add a second dimension: the Pro plan allows 30,000 requests per five-hour window and 300,000 per week on agnes-2.0-flash, while video generation is capped at 500 seconds of output per day across all paid subscription tiers. These numbers are described as operational reference values as of late June 2026, not permanent guarantees — a candid disclaimer that most API providers bury.

For agent builders, the OpenAI compatibility is the practical entry point. The Python quick-start literally imports openai.OpenAI, sets a custom base_url, and proceeds identically to any OpenAI workflow. Tool calling and thinking mode are listed as supported on compatible chat endpoints, though the README flags that you should verify model-specific docs before enabling advanced parameters in production — reasonable caution.

What this repo actually is, stripped of the marketing framing, is a developer onboarding document and model reference sheet. There are no model weights, no training details, no benchmark methodology beyond a mention of something called PinchBench. The claim that models are trained in-house is asserted but not substantiated anywhere in the README. Whether the underlying models are genuinely proprietary or wrappers around other providers' infrastructure is not addressed. That gap matters if you're evaluating this for production use rather than prototyping.