Agnes AI: OpenAI-compatible API with video generation and a free tier

What it does This repository is the official documentation hub and model catalog for Agnes AI, a hosted platform offering OpenAI-compatible endpoints for text, image, video, and agent workflows. It publishes rate limits, quota tiers, and bilingual troubleshooting guides alongside minimal client examples. In practice, it is glue and reference material for a proprietary API rather than open-source model weights.

The interesting bit The platform bets on drop-in SDK compatibility: you can point an existing openai client at apihub.agnes-ai.com/v1 and start calling multimodal models, including an async video generator that returns a video_id for polling. The README even notes that agnes-2.0-flash had its context window rolled back from a temporary 1M to 256K tokens in June 2026, an operational detail many providers prefer to bury.

Key highlights

• agnes-2.5-flash offers a 512K context window and tool-calling support via standard /v1/chat/completions
• agnes-image-2.1-flash and agnes-video-v2.0 extend the same OpenAI-shaped client to image and video generation
• Free tier grants 20 RPM on text models and 1 RPM on video; the Token Plan jumps to 1,000 RPM text and 500 seconds of video per day
• Bilingual documentation (English and Simplified Chinese) includes explicit error-code guides and per-plan quota tables
• Async video tasks require video_id polling; the docs explicitly warn against using legacy task_id queries

Caveats

• The repository contains no model weights or inference code; it is purely documentation and examples for a proprietary, hosted service
• Context windows and rate limits are explicitly labeled “operational limits, not permanent guarantees” and have already changed once in June 2026
• Video generation quotas remain modest even on paid tiers (500 seconds per day on the Token Plan)

Verdict Worth a look if you need an OpenAI-compatible alternative with video generation and transparent quota tables, but pass if you are hunting for open-source weights or self-hosted inference.