Every model, with its real price.

from openai import OpenAI

client = OpenAI(
    base_url="https://api.inferall.ai/v1",
    api_key="ifu_your_inferall_key",
)

resp = client.chat.completions.create(
    model="nvidia/nemotron-3-ultra-550b-a55b",   # any id from the free table below
    messages=[{"role": "user", "content": "Hello"}],
)

Guides for popular free models

• gpt-oss-20b (openai/gpt-oss-20b)
• Nemotron 3 Ultra (nvidia/nemotron-3-ultra-550b-a55b)
• Nemotron 3.5 Super VL (nvidia/nemotron-3.5-super-vl-120b-a12b)
• Step 5 Preview (stepfun-ai/step-5-preview)
• Nemotron 3 Nano Omni Reasoning (nvidia/nemotron-3-nano-omni-30b-a3b-reasoning)
• DiffusionGemma 26B (google/diffusiongemma-26b-a4b-it)
• Nemotron 3.5 Lightning (nvidia/nemotron-3.5-lightning-30b-a3b)

Free open models (15)

$0 input and $0 output, served on NVIDIA NIM. A new trial key can call these without a card. If yours returns a 402 here, your account already started checkout at some point and needs it finished before free calls resume.

Uncensored models (2)

Chat models without the refusal layer most providers add. Same API and key as everything else here; send the ID exactly as written. Billed from your balance, so a trial key needs the $5 starter pack first. inferall-uncensored has a 262k context window; inferall-uncensored-large has 1M. Reasoning is off by default; pass reasoning_effort to turn it on. You are responsible for how you use the output. Guide and examples.

Premium text models (59)

Billed from your balance at the rates shown, cheapest input first: the provider's published rate plus 30%, or 15% below it on Anthropic models, with no per-request minimum. The provider's own rate is in the last column. Some models cost more for a long prompt: OpenAI's GPT-5.4, GPT-5.5, GPT-5.6 and GPT-6 models above 272K input tokens, Gemini Pro and Grok above 200K. Those calls bill at the provider's long-context rate, plus the same margin. A trial key cannot call these until you add a card. The IDs below already include the provider prefix they need (openai/, anthropic/, or gemini/). Our catalog stores them bare, because that is the key we price against once the prefix has been read off. Sent that way, as gpt-4o rather than openai/gpt-4o, they route to the free NVIDIA tier. The reply does say so: its model field names the model that actually answered, so compare it against the ID you sent. What you will not get is an error or a warning header, so code that never reads response.model will not notice.

Image, video, and audio (37)

Billed per image or per second of output rather than per token, at the price shown: the provider's rate plus 30%, with a minimum of $0.01 per image or $0.005 per second.