A multimodal API compatible with the OpenAI standard, with its main models available for free. Here’s what’s real, what the limits are, and what the risks are.
Agnes AI, from Sapiens AI (Singapore), offers its own text, vision, image, and video models through an OpenAI-compatible API. “Free” means no charge per token or generation — not unrestricted or guaranteed use, or suitability for critical production.
Text, images, and video currently at US$ 0 — including via API. One of the most aggressive free offers of 2026.
Same call format as OpenAI SDKs: integrate with n8n, Python, Node, agents, and existing apps by changing only the base URL and key.
No uptime guarantee, limits may change, and on the free plan your data may be used to train models (unless you opt out). Don’t use it as your only provider in a critical system.
A single key and a single gateway (apihub.agnes-ai.com) for text/vision, images, and video.
Text, vision, and agents. Conversation, reasoning, code, tool calls, streaming, and vision via image URL. Declared context of 512K tokens (external catalogs cite 256K — test before sending very large contexts). US$ 0 per million tokens.
POST https://apihub.agnes-ai.com/v1/chat/completions model: agnes-2.0-flash
Text→image, image→image, redesign, instruction-based editing, returns as URL or Base64, resolutions from 1K to 4K in 1:1, 16:9, 9:16, 4:3, 3:4, 2:3, 3:2, and 21:9 aspect ratios. US$ 0 per image.
POST https://apihub.agnes-ai.com/v1/images/generations model: agnes-image-2.1-flash
Text→video, image→video, keyframes, prompt-based camera/motion control, 480p/720p/1080p, asynchronous generation. Up to 441 frames per request (formula 8n+1). US$ 0 per second.
POST https://apihub.agnes-ai.com/v1/videos GET .../agnesapi?video_id=ID
| Aspect ratio | 1K | 2K | 3K | 4K |
|---|---|---|---|---|
| 1:1 | 1024×1024 | 2048×2048 | 3072×3072 | 4096×4096 |
| 16:9 | 1312×736 | 2624×1472 | 3936×2208 | 5248×2944 |
| 9:16 | 736×1312 | 1472×2624 | 2208×3936 | 2944×5248 |
| 4:3 | 1152×864 | 2304×1728 | 3456×2592 | 4608×3456 |
| 3:4 | 864×1152 | 1728×2304 | 2592×3456 | 3456×4608 |
Dimensions outside the table (e.g., 1920×1080) may be converted automatically. For a YouTube thumbnail: request "size":"2K","ratio":"16:9" and resize to 1920×1080.
Because the API is OpenAI-compatible, the usual approach works—and the safest architecture uses Agnes as a free, high-volume layer, never as the only brain.
Free, high-volume tasks: batch thumbnails, summaries, classification, test agents, experimental videos.
Gemini / GLM / DeepSeek as an alternative route when Agnes is unavailable or limited.
OpenAI / Claude for workloads that require predictable quality, an SLA, and broad validation.
The free plan is advertised as free indefinitely, but without a contractual guarantee: availability, quotas, and policies may change. Paid plans mainly buy speed and priority, not “unlimited use.”
| Feature | Effective limit |
|---|---|
| Text | 20 req/min |
| 1K images | 20 img/min |
| 2K images | 10 img/min |
| 3K / 4K images | 1 img/min |
| Video | 1 req/min |
Theoretically 28.800 text calls/day and 1.200 1K images/hour — but queues, fair use, and capacity reduce actual volume. No SLA.
| Plan | Price/month* | Text / 5h | Text / week |
|---|---|---|---|
| Starter | US$ 4 | 1.500 | 15.000 |
| Plus | US$ 10 | 7.500 | 75.000 |
| Pro | US$ 50 | 30.000 | 300.000 |
All Token Plans: up to 1.000 RPM for text, 4.000 images/day, 500 s of video/day, 100 RPM (1K) / 80 RPM (2K) for images, and 5 RPM for video. *Prices found in a community source — confirm at checkout. Odd detail: the paid plan increases speed but introduces explicit time-based quotas that aren’t published for the free plan.
You can combine: a free key and a Token Plan key use separate pools — when the paid quota runs out, the free key still works within the free limits. Creating multiple keys in the same category no multiplies limits (they share the same pool).
Everything through the OpenAI-compatible endpoint. Change $AGNES_API_KEY with your key.
Sign up through the official campaign — agnes-ai.com/campaign — and generate a free key. Save it as an environment variable — never in the code.
# access: https://agnes-ai.com/campaign export AGNES_API_KEY="sua-chave-aqui"
OpenAI-format chat completions, with streaming and tool calls.
curl https://apihub.agnes-ai.com/v1/chat/completions \ -H "Authorization: Bearer $AGNES_API_KEY" \ -H "Content-Type: application/json" \ -d '{"model":"agnes-2.0-flash","messages":[{"role":"user","content":"Resuma o que é a Agnes AI em 3 frases."}]}'
In Python (or Node), just change base_url — the rest of the code doesn’t change.
# pip install openai from openai import OpenAI client = OpenAI(base_url="https://apihub.agnes-ai.com/v1", api_key="$AGNES_API_KEY") r = client.chat.completions.create(model="agnes-2.0-flash", messages=[{"role":"user","content":"Olá!"}]) print(r.choices[0].message.content)
E.g., 16:9 thumbnail in 2K (then resize to 1920×1080).
curl https://apihub.agnes-ai.com/v1/images/generations \ -H "Authorization: Bearer $AGNES_API_KEY" \ -H "Content-Type: application/json" \ -d '{"model":"agnes-image-2.1-flash","prompt":"thumbnail de YouTube sobre IA gratuita, estilo tech dark","size":"2K","ratio":"16:9"}'
Create the job, receive a video_id and poll until it’s ready. Frames follow the formula 8n+1 (max. 441).
# create curl https://apihub.agnes-ai.com/v1/videos \ -H "Authorization: Bearer $AGNES_API_KEY" \ -H "Content-Type: application/json" \ -d '{"model":"agnes-video-v2.0","prompt":"drone sobrevoando cidade futurista ao entardecer","resolution":"720p"}' # check the result curl "https://apihub.agnes-ai.com/agnesapi?video_id=SEU_ID" \ -H "Authorization: Bearer $AGNES_API_KEY"
On the free plan, prompts and files may be used to train models unless you opt out—available within the service or through support. Never send sensitive data (CPF, patients, contracts, keys, LGPD data) through the free plan.
# support channel for opting out of training use
support@agnes-ai.com
This isn’t theory: it came from ~70 real API calls, comparing what the documentation promises with what you have observed. The full record — accepted parameters, those that fail with HTTP 400, quota, cost, and open questions — is in NOTAS-API.md.
It’s not a matter of quality (they’re tied)—it’s the content filter: in Portuguese, it blocks legitimate generation with HTTP 400. Write the prompt in English, even for content in PT.
Any animal with a tail in a front-facing pose comes out with two. The fix is to be explicit in the prompt: ONE SINGLE bushy tail. This applies to the general default — whatever you don’t specify, the model duplicates.
About 34% of calls fail with 503 — and retries recover nearly 100%. Without retries, you may wrongly conclude that the API “doesn’t work.”
Descriptors such as fur, expressive eyes or children's book inject characters in a prompt that was only meant to describe a setting. In the style description, use aesthetic terms only.
1K takes ~32s; 4K takes ~153s and fails much more often. Resolution costs nothing extra (it’s all US$ 0), but costs time and increases the error rate — generate at 1K when volume matters.
The return URL is temporary. Save the file in the same step as generation; don’t keep the link assuming you can retrieve it later.
response_format at the JSON root level returns HTTP 400 — it only works inside extra_body. Worse: unknown parameters are silently discarded (the gateway uses drop_params), so a 200 OK no proves that your parameter was used. And there is no seed or fine-tuning/LoRA.
curl https://apihub.agnes-ai.com/v1/images/generations \ -H "Authorization: Bearer $AGNES_API_KEY" -H "Content-Type: application/json" \ -d '{"model":"agnes-image-2.1-flash","prompt":"...","size":"1K","ratio":"16:9", "extra_body":{"response_format":"url"}}'
The evidence is strong for agents and reasonable for images; in video, the model is far from the leaders. Broad, independent text benchmarks (SWE-bench, GPQA, MMLU-Pro, etc.) are lacking.
Claw-Eval: Pass³ 60.9% — 9th place (snapshot leader: Claude Opus 4.6 at 70.4%). A legitimate benchmark of 300 real tasks verified by humans. Competitive, but not the leader — and a single benchmark does not prove overall superiority in coding, Portuguese, math, or factuality.
~1.178–1.184 Elo in editing evaluations: capable and competitive. Great for thumbnail alternatives, banners, backgrounds, and A/B testing; for premium assets (hands, typography, exact logos), review them or finish in a stronger model.
Ranked around 61 (T2V) and 63 (I2V) — behind the leaders in realism and complex motion. Before it was free, it cost about US$ 0.30/min through gateways. At US$ 0, it’s worth testing extensively for supporting scenes and high volume.
The offer is too compelling to ignore for testing, content, agents, and non-critical automation. But don’t use it as the sole provider for a critical commercial system.
Aplicação ↓ Roteador de modelos ├── Agnes AI: tarefas gratuitas e alto volume ├── Gemini/GLM/DeepSeek: segunda opção └── OpenAI/Claude: tarefas críticas ou complexas