AI API Router Directory

Groq runs LPU inference hardware tuned for raw tokens-per-second. The API offers a documented free developer tier — per-model rate limits such as 30 RPM / 1,000 RPD / 8K TPM on gpt-oss-120b — with full x-ratelimit-* header observability, and paid Developer plans that add Batch and Flex processing plus higher ceilings.

Community reported approximate monthly value of free tier rate limits for individual developers.

Fireworks AI is a self-serve inference platform: per-token serverless (Standard/Priority/Fast tiers) with $1 in free credits and postpaid billing, fine-tuning (LoRA and full-param, priced per 1M training tokens), a shared-pool Serverless Training API, and on-demand GPUs from $8/hr (H100, from Sep 1 2026) to $20/hr (GB300).

Official pricing page: 'Get started with $1 in free credits' (confirmed 2026-09-16).

Hyperbolic (hyperbolic.xyz) is the API-specific endpoint for Hyperbolic's AI cloud. It provides serverless inference for over 25 open-source models through an OpenAI-compatible /v1 endpoint. The platform supports text, image, and audio generation with pay-as-you-go pricing.

~$1 trial credit for new users (DeepSeek V3, Llama 3.1 405B, Qwen); pay-as-you-go after it is used (Llama 405B ~$4/M, DeepSeek R1 ~$3/M tokens).

Nebius AI Studio is a European-based GPU cloud platform providing managed inference for top open-source LLMs like Llama 3.1 and DeepSeek R1. It leverages its own EU data centers to offer high-performance, privacy-compliant inference with transparent pay-as-you-go pricing tailored for both developers and enterprises.

AI Studio grants new accounts ~$1 one-time trial credit (startup/promo programs can grant far more); Nebius AI Cloud free-trial program suspended as of 2026-07-13; studio models otherwise pay-as-you-go.

Scaleway Generative APIs is a serverless, OpenAI-compatible inference service hosted only in Paris and billed per 1M tokens in euros. Published rates run from EUR 0.15/1M input and EUR 0.60/1M output for gpt-oss-120b, and EUR 0.15/EUR 0.35 for mistral-small-3.2-24b-instruct, up to EUR 1.80/EUR 5.50 for glm-5.2; whisper-large-v3 is metered per audio minute at EUR 0.003 with free output tokens, and embedding models are input-only at EUR 0.10/1M. Every new customer gets a free tier of 1,000,000 tokens, with billing starting at token 1,000,001.

1,000,000 free tokens per new customer, stated on the Generative APIs page; valued at the cheapest published input rate of EUR 0.15/1M (mistral-small-3.2-24b-instruct or gpt-oss-120b), which is roughly USD 0.16 - the grant is token-denominated, so actual dollar value varies by model.

Cloudflare Workers AI serves open-weight models on Cloudflare's edge network, metered in Neurons ($0.011 per 1,000). Every account — including the Free plan — gets 10,000 Neurons per day free, resetting at 00:00 UTC; beyond that, the Paid plan bills usage. The catalog spans LLMs, embeddings, image and audio models and sits next to your Workers apps.

10,000 Neurons/day free on all plans (reset 00:00 UTC) ≈ $0.11/day ≈ $3.30/month at full use; overage $0.011 per 1,000 Neurons on the Paid plan.

Hugging Face Inference Providers routes one API key across 200+ models from multiple inference providers at the providers' own rates — no markup. Every user gets monthly credits ($0.10 on the Free plan; $2.00 PRO; $2.00/seat Team & Enterprise) applied automatically before pay-as-you-go; a BYOK mode routes through your own provider keys instead.

Official docs: $0.10/month Inference-Providers credits for Free users (PRO $2.00/month, Team/Enterprise $2.00/seat), credited monthly and applied before pay-as-you-go (confirmed 2026-09-16).

Hyperbolic (hyperbolic.ai) is an AI cloud that sells on-demand, reserved and private GPU capacity and, separately, a serverless inference API described as OpenAI-compatible at base URL https://api.hyperbolic.xyz/v1 with 25+ open-source models. Its official rate-limit table defines a Basic tier of 60 requests per minute available to a free account with a $0 minimum deposit, and a Pro tier of 600 requests per minute that unlocks automatically once $5 or more is deposited. Serverless inference is pay-as-you-go: text generation runs from $0.10 per 1M tokens for small models, $0.20-$0.40 for 32B-72B models and $0.30-$4.00 for 120B-480B models, with images at a $0.01 base rate per 1024x1024 image at 25 steps and audio at $5.00 per 1M characters.

No free tier stated on official pages (checked 2026-09-16): Hyperbolic documents a free "Basic" tier as 60 requests/minute with a Free account, but no dollar credit or trial allowance is published on the pricing or billing pages.

Novita AI runs 200+ models across text, image, audio and video behind an OpenAI-compatible API at api.novita.ai/openai, plus serverless endpoints, GPU instances and agent sandboxes, all billed per token or per GPU-second. Published model rates start at $0.02/1M input and $0.05/1M output for Llama 3.1 8B Instruct, with DeepSeek V4 Flash at $0.14/$0.28 on a 1M-token context, and three Ling 3.0 Flash variants (Fin, VL, Sante) listed at Free input and Free output with 256K context. Batch inference carries an introductory 50% discount on input and output tokens for supported models.

No standing free credit is stated on the current official pricing or docs pages (checked 2026-09-16); the pricing table lists Ling 3.0 Flash Fin/VL/Sante at Free input and output (256K context), and an official blog post from May 2025 advertised a limited-time $10 new-user credit that current pages do not repeat.