llms.txt
Compare with other sites' llms.txt files →# Runcrate
> Runcrate is an AI cloud platform combining an OpenAI-compatible inference API for open and frontier models with on-demand GPU rentals (H100, H200, B200, A100, L40S, RTX 5090, RTX 4090, and more) and reserved clusters on request. GPUs rotate in and out of stock daily: https://www.runcrate.ai/api/gpu-pricing lists what is in stock now. One API key, one bill, one dashboard for chat completion, image / video generation, speech, transcription, plus dedicated GPU instances, persistent volumes, fine-tuning, and reserved clusters.
Runcrate is self-serve. Inference billing is per-token / per-image / per-second / per-minute (varies by model). GPU rentals bill by the second on most GPUs and by the minute on a few. There are no commitments, no minimums, and credits never expire. The inference API is a drop-in replacement for the OpenAI Python and TypeScript SDKs: set `base_url=https://api.runcrate.ai/v1` and existing code keeps working.
## Get started
- [Homepage](https://www.runcrate.ai/): Top-level platform overview.
- [Sign up](https://www.runcrate.ai/auth/signup): Create an account, get an API key, add credits.
- [Sign in](https://www.runcrate.ai/auth/login): Log in to dashboard.
- [Dashboard](https://www.runcrate.ai/dashboard): Console for instances, models, API keys, billing.
- [Blog](https://www.runcrate.ai/blog): What shipped on Runcrate and how to use it, with the docs and commands for each change. Feed: https://www.runcrate.ai/blog/rss.xml
- [Releases](https://www.runcrate.ai/blog/category/releases): Roundups of what changed across Runcrate, grouped by what you can now do.
- [Clusters](https://www.runcrate.ai/blog/category/clusters): Starting, connecting and paying for several GPU machines on one private network.
- [Compute](https://www.runcrate.ai/blog/category/compute): Running work on a GPU: machines you keep for as long as you need, and jobs that stop on their own.
- [Inference](https://www.runcrate.ai/blog/category/inference): Calling open models, and seeing how long each call took, what it cost and why it stopped.
<!-- generated:connect-an-agent:start -->
## Connect an agent
- [MCP server](https://mcp.runcrate.ai/mcp): streamable HTTP; the client signs in through the browser (OAuth with dynamic registration), with no API key. Steps for any coding agent to add it to itself: https://runcrate.ai/mcp.md. The line a person pastes into their agent: `Add Runcrate's MCP server to yourself: run curl -fsSL "https://runcrate.ai/mcp.md" and follow it.`
- [Add a Runcrate model to a project](https://runcrate.ai/agents.md?intent=build): steps for a coding agent (SDK, key in the env file, one function, one test call). Add `model=<model id>` for one model. The line: `Add a Runcrate model to this project: run curl -fsSL "https://runcrate.ai/agents.md?intent=build" and follow it.`
- [Run a coding agent on a Runcrate model](https://runcrate.ai/agents.md?intent=agent): Claude Code, Codex, OpenCode, Cline, Roo Code, Kilo Code and Cursor. The line: `Set yourself up to use a Runcrate model: run curl -fsSL "https://runcrate.ai/agents.md?intent=agent" and follow it.`
- Base URLs: https://api.runcrate.ai/v1 (OpenAI-compatible), https://api.runcrate.ai (Anthropic Messages, for Claude Code). The key goes in `RUNCRATE_API_KEY`; make one at https://www.runcrate.ai/dashboard/api-keys.
- [Use Runcrate from your agent](https://www.runcrate.ai/docs/getting-started/agents): one-line MCP installs for Claude Code (`claude mcp add --transport http --scope user runcrate https://mcp.runcrate.ai/mcp`), Codex (`codex mcp add runcrate --url https://mcp.runcrate.ai/mcp`), Gemini CLI (`gemini mcp add --transport http --scope user runcrate https://mcp.runcrate.ai/mcp`) and Cursor (https://cursor.com/en/install-mcp?name=runcrate&config=eyJ1cmwiOiJodHRwczovL21jcC5ydW5jcmF0ZS5haS9tY3AifQ==).
- Prices for agents, read live with no key: [pricing.md](https://www.runcrate.ai/pricing.md) (every GPU in stock per GPU-hour on demand, and every model per 1M tokens or per unit, with the read time) and [gpus.json](https://www.runcrate.ai/gpus.json) (every GPU: memory, price, stock, regions, prepaid terms).
- [Agent skill](https://github.com/runcrateai/runcrate-mcp/tree/main/skills/runcrate) (Agent Skills format): check prices, launch a job or instance with a max runtime, stop it, top up. Install: `npx skills add runcrateai/runcrate-mcp --skill runcrate`
- [CLI](https://www.runcrate.ai/docs/cli/overview): `curl -sSL https://runcrate.ai/install.sh | sh` (macOS, Linux), `brew install runcrateai/tap/runcrate` or `irm https://runcrate.ai/install.ps1 | iex` (Windows), then `runcrate login`.
- [Docs](https://www.runcrate.ai/docs): guides and the API reference.
<!-- generated:connect-an-agent:end -->
## Pricing
- [GPU pricing](https://www.runcrate.ai/pricing): Live on-demand GPU rental rates for every GPU in stock (current prices on the page, as markdown with every model's price at https://www.runcrate.ai/pricing.md, and as JSON at https://www.runcrate.ai/gpus.json). Billed by the second on most GPUs, by the minute on a few. No quota requests.
- [Inference API pricing](https://www.runcrate.ai/inference/pricing): Per-token / per-image / per-second / per-minute pricing for every active model, with the live count on the page. Server-rendered, includes a JSON-LD ItemList of every model with prices in the page source: readable by LLM agents without JS.
- [On-demand GPU compute](https://www.runcrate.ai/compute): On-demand GPU rental product page.
- [Storage volumes](https://www.runcrate.ai/storage): Persistent volumes at $0.03 per GB a month, prepaid weekly, 10 GB to 10,000 GB. Files are on the instance's own disk at /workspace, read and written at local-disk speed; attach to any GPU in any region at launch (storage_id), one running instance at a time.
- [Reserved GPU clusters](https://www.runcrate.ai/clusters/reserved): Reserve a multi-node GPU cluster for a month or longer, including GPUs that are not on demand. Describe it once; Runcrate verifies your company and your need first (a mutual NDA, your company details, proof of funds, and what you'll run and from when), then goes to the GPU owners that match your exact spec. They quote in sealed bids, every offer arrives side by side in one table, and you sign directly with the owner you pick. A person is available at any step.
## Benchmarks
Measured runs on Runcrate GPUs. Each result shows its run date (single runs), hardware, software and method, the run-day cost, and the cost at today's live price per GPU-hour.
- [Benchmarks](https://www.runcrate.ai/benchmarks): Every benchmark page and every run, newest first: measured runs on rented GPUs, each with its run date, hardware, software versions, method and cost.
- [LLM inference benchmarks](https://www.runcrate.ai/benchmarks/llm-inference): Qwen2.5-7B on vLLM on RTX A6000, L40S and H100: output tokens per second, time to first token and cost per 1M output tokens; FP8 vs bf16 on H100; prefix caching on L40S; Ollama Q4/Q8 vs vLLM bf16 on an A10.
- [A100 vs H100 vs H200 benchmark](https://www.runcrate.ai/benchmarks/a100-vs-h100-vs-h200): One script on H200, H100, A100 80GB, L40S, RTX A6000 and RTX A4000: dense matmul TFLOPS by data type, memory copy bandwidth, nanoGPT, a LoRA fine-tune and SDXL, with the compute cost of each.
- [Whisper benchmarks](https://www.runcrate.ai/benchmarks/whisper): Whisper large-v3 and large-v3-turbo on RTX 5090, L40S, RTX A6000 and H100, five setups: speed against real time, word error rate and cost per 1,000 audio hours.
- [Embedding benchmarks](https://www.runcrate.ai/benchmarks/embeddings): BGE small, base and large at 128 and 512 tokens on RTX 5090, L40S, RTX A6000 and H100: passages per second and cost per 1M passages.
- [Training and fine-tuning benchmarks](https://www.runcrate.ai/benchmarks/training): Full fine-tune vs LoRA vs QLoRA on an L40S, PyTorch DDP on 4 PCIe RTX A6000s, and torch.compile vs eager on four GPUs: memory, speed and training cost.
- [SDXL benchmarks](https://www.runcrate.ai/benchmarks/sdxl): SDXL base at 1024 x 1024, 30 steps, on RTX A6000, L40S, RTX 5090 and H100: seconds per image eager and compiled, and cost per 1,000 images.
- [GPU startup time](https://www.runcrate.ai/benchmarks/gpu-startup-time): Seconds from the create call to the first CUDA matmul on five GPU types, and a 7B model start from a fresh Hugging Face download against a local cache on an L40S.
## GPU pricing pages
Per-GPU rental detail with cross-cloud pricing comparison, workloads, benchmarks, and full specs. URL pattern: `/pricing/gpu/{id}`. Each page shows a live price and on-demand deploy only when that GPU is in stock; a GPU with no stock reads "Not in stock right now" and offers reserved capacity or a register-interest form plus the closest in-stock GPU. Stock and prices change daily: read the page or https://www.runcrate.ai/api/gpu-pricing (every GPU in stock now, with its cheapest hourly rate per GPU; a GPU sold out within 30 days stays listed with `available_options` 0, its last price and `last_in_stock_at`). Each GPU in stock also lists `prepaid_terms`: for each prepaid length its machines offer now below the GPU's on-demand `min_price_hr` (1, 3, 6 or 12 months, paid upfront, can't be cancelled), `months`, `price_per_gpu_hour`, `discount_pct` (a fraction of that machine's on-demand rate: 0.05 is 5%), `on_demand_price_per_gpu_hour` (that machine's on-demand rate per GPU), `total_per_gpu` for the whole term and `num_gpus` (the GPUs in that machine, all paid for); an empty list when none is offered.
- [B200 pricing](https://www.runcrate.ai/pricing/gpu/b200): NVIDIA B200, 180GB HBM3e: live stock and price on the page; reserved clusters on request
- [B300 pricing](https://www.runcrate.ai/pricing/gpu/b300): NVIDIA B300 Blackwell Ultra, 288GB HBM3e per GPU: live stock and price on the page; reserved clusters on request
- [H200 pricing](https://www.runcrate.ai/pricing/gpu/h200): NVIDIA H200 SXM, 141GB HBM3e: live stock and price on the page
- [H200 NVL pricing](https://www.runcrate.ai/pricing/gpu/h200-nvl): NVIDIA H200 NVL, the PCIe card on NVLink bridges, 141GB HBM3e: live stock and price on the page
- [H100 pricing](https://www.runcrate.ai/pricing/gpu/h100): NVIDIA H100, 80GB HBM3: live stock and price on the page
- [GH200 pricing](https://www.runcrate.ai/pricing/gpu/gh200): NVIDIA GH200 Grace Hopper superchip, 96GB HBM3: live stock and price on the page
- [A100 80GB pricing](https://www.runcrate.ai/pricing/gpu/a100-80gb): NVIDIA A100 80GB, 80GB HBM2e: live stock and price on the page
- [A100 40GB pricing](https://www.runcrate.ai/pricing/gpu/a100-40gb): NVIDIA A100 40GB, 40GB HBM2: live stock and price on the page
- [GB200 pricing](https://www.runcrate.ai/pricing/gpu/gb200): NVIDIA GB200 Grace+Blackwell superchip, 186GB HBM3e per GPU (372GB per superchip): live stock on the page; reserved clusters on request
- [GB300 pricing](https://www.runcrate.ai/pricing/gpu/gb300): NVIDIA GB300 Blackwell Ultra, 288GB HBM3e per GPU: live stock on the page; reserved clusters on request
- [RTX 5090 pricing](https://www.runcrate.ai/pricing/gpu/rtx-5090): NVIDIA RTX 5090, 32GB GDDR7: live stock and price on the page
- [RTX 4090 pricing](https://www.runcrate.ai/pricing/gpu/rtx-4090): NVIDIA RTX 4090, 24GB GDDR6X: live stock and price on the page
- [RTX 3090 pricing](https://www.runcrate.ai/pricing/gpu/rtx-3090): NVIDIA RTX 3090, 24GB GDDR6X: live stock and price on the page
- [L40S pricing](https://www.runcrate.ai/pricing/gpu/l40s): NVIDIA L40S, 48GB GDDR6: live stock and price on the page
- [A40 pricing](https://www.runcrate.ai/pricing/gpu/a40): NVIDIA A40, 48GB GDDR6: live stock and price on the page
- [RTX A6000 pricing](https://www.runcrate.ai/pricing/gpu/rtx-a6000): NVIDIA RTX A6000, 48GB GDDR6: live stock and price on the page
- [V100 pricing](https://www.runcrate.ai/pricing/gpu/v100): NVIDIA Tesla V100, 32GB HBM2: live stock and price on the page
- [RTX 4080 pricing](https://www.runcrate.ai/pricing/gpu/rtx-4080): NVIDIA RTX 4080, 16GB GDDR6X: live stock and price on the page
- [RTX 4070 Ti pricing](https://www.runcrate.ai/pricing/gpu/rtx-4070-ti): NVIDIA RTX 4070 Ti: live stock and price on the page
- [MI300X pricing](https://www.runcrate.ai/pricing/gpu/mi300x): AMD MI300X, 192GB HBM3: live stock and price on the page; reserved clusters on request
- [MI250X pricing](https://www.runcrate.ai/pricing/gpu/mi250x): AMD MI250X, 128GB HBM2e: live stock and price on the page
- [MI210 pricing](https://www.runcrate.ai/pricing/gpu/mi210): AMD MI210: live stock and price on the page
## GPU model spec pages
Spec deep-dive (VRAM, bandwidth, FP16/FP32, TDP, architecture). Same id list as above.
- [/gpu-models/b200](https://www.runcrate.ai/gpu-models/b200): H100 / H200 / B200 / B300 / GB200 / GB300 / MI300X model pages all follow the same `/gpu-models/{id}` pattern.
- [/gpu-models/b300](https://www.runcrate.ai/gpu-models/b300)
- [/gpu-models/h200](https://www.runcrate.ai/gpu-models/h200)
- [/gpu-models/h200-nvl](https://www.runcrate.ai/gpu-models/h200-nvl)
- [/gpu-models/h100](https://www.runcrate.ai/gpu-models/h100)
- [/gpu-models/gh200](https://www.runcrate.ai/gpu-models/gh200)
- [/gpu-models/a100-80gb](https://www.runcrate.ai/gpu-models/a100-80gb)
- [/gpu-models/a100-40gb](https://www.runcrate.ai/gpu-models/a100-40gb)
- [/gpu-models/gb200](https://www.runcrate.ai/gpu-models/gb200)
- [/gpu-models/gb300](https://www.runcrate.ai/gpu-models/gb300)
- [/gpu-models/rtx-5090](https://www.runcrate.ai/gpu-models/rtx-5090)
- [/gpu-models/rtx-4090](https://www.runcrate.ai/gpu-models/rtx-4090)
- [/gpu-models/rtx-3090](https://www.runcrate.ai/gpu-models/rtx-3090)
- [/gpu-models/l40s](https://www.runcrate.ai/gpu-models/l40s)
- [/gpu-models/a40](https://www.runcrate.ai/gpu-models/a40)
- [/gpu-models/rtx-a6000](https://www.runcrate.ai/gpu-models/rtx-a6000)
- [/gpu-models/v100](https://www.runcrate.ai/gpu-models/v100)
- [/gpu-models/mi300x](https://www.runcrate.ai/gpu-models/mi300x)
- [/gpu-models/mi250x](https://www.runcrate.ai/gpu-models/mi250x)
- [/gpu-models/mi210](https://www.runcrate.ai/gpu-models/mi210)
- [/gpu-models/rtx-4080](https://www.runcrate.ai/gpu-models/rtx-4080)
- [/gpu-models/rtx-4070-ti](https://www.runcrate.ai/gpu-models/rtx-4070-ti)
## GPU comparisons
Side-by-side comparison pages with TL;DR verdicts, bar charts, and full spec tables.
- [GPU comparisons index](https://www.runcrate.ai/gpu-compare): Browse all comparisons grouped by architecture.
- [H100 vs A100 80GB](https://www.runcrate.ai/gpu-compare/h100-vs-a100-80gb)
- [B200 vs H200](https://www.runcrate.ai/gpu-compare/b200-vs-h200)
- [B200 vs B300](https://www.runcrate.ai/gpu-compare/b200-vs-b300)
- [H200 vs MI300X](https://www.runcrate.ai/gpu-compare/h200-vs-mi300x)
- [GB200 vs GB300](https://www.runcrate.ai/gpu-compare/gb200-vs-gb300)
- [H100 vs RTX 5090](https://www.runcrate.ai/gpu-compare/h100-vs-rtx-5090)
- [RTX 5090 vs RTX 4090](https://www.runcrate.ai/gpu-compare/rtx-5090-vs-rtx-4090)
- [H200 vs RTX 5090](https://www.runcrate.ai/gpu-compare/h200-vs-rtx-5090)
- URL pattern: `/gpu-compare/{a}-vs-{b}`: works for any pair of GPU ids in the spec list above.
## Model catalog
- [Inference catalog (HTML)](https://www.runcrate.ai/inference/pricing): Searchable, category-filtered model catalog with live pricing.
- [Model JSON catalog](https://www.runcrate.ai/api/models/catalog): Machine-readable JSON of every active model: id, display_name, prices, parameters, supported_parameters, context_length, category. Prices are the exact rates each request is billed.
## API landing pages: by capability
Topic-specific pages with code examples and feature comparisons. Each is OpenAI-compatible.
- [APIs index](https://www.runcrate.ai/apis): Every API page, the platform pages and the GPU-cloud comparisons, in one list.
- [Speech-to-text API](https://www.runcrate.ai/speech-to-text-api): Whisper / Voxtral, fast, accurate transcription.
- [Real-time STT](https://www.runcrate.ai/real-time-stt): Low-latency streaming transcription.
- [Whisper API](https://www.runcrate.ai/whisper-api): Dedicated OpenAI Whisper Large v3 inference.
- [Text-to-speech API](https://www.runcrate.ai/text-to-speech-api): Qwen3 TTS, Kokoro, Orpheus.
- [DeepSeek API](https://www.runcrate.ai/deepseek-api): DeepSeek V3.2 and R1 inference.
- [Mistral API](https://www.runcrate.ai/mistral-api): Mistral Small, Nemo, Mixtral and Voxtral.
- [Llama API](https://www.runcrate.ai/llama-api): Meta Llama 4 and Llama 3.
- [Qwen API](https://www.runcrate.ai/qwen-api): Qwen3 chat, vision, TTS, code.
- [Gemma API](https://www.runcrate.ai/gemma-api): Google Gemma model inference.
- [Ollama Cloud](https://www.runcrate.ai/ollama-cloud): Run Ollama-compatible models in the cloud.
- [Stable Diffusion API](https://www.runcrate.ai/stable-diffusion-api): Stable Diffusion 3.5, Juggernaut XL and FLUX image generation.
- [FLUX API](https://www.runcrate.ai/flux-api): FLUX.1 Dev / Schnell / Pro.
- [Image generation API](https://www.runcrate.ai/image-generation-api): FLUX, Imagen, Stable Diffusion, Ideogram.
- [Veo API](https://www.runcrate.ai/veo-api): Google Veo 3 / Veo 2 video.
- [Kling API](https://www.runcrate.ai/kling-api): Kling video generation.
- [RunwayML API alternative](https://www.runcrate.ai/runway-api): Multi-provider video generation.
- [Chat API](https://www.runcrate.ai/chat-api): Multi-provider chat completions.
- [Embedding API](https://www.runcrate.ai/embedding-api): Vector embeddings for RAG and search.
- [Hugging Face Inference alternative](https://www.runcrate.ai/huggingface-inference): Open models via an OpenAI-compatible API.
- [Serverless inference](https://www.runcrate.ai/serverless-inference): GPU inference without managing infrastructure.
- [AI agents API](https://www.runcrate.ai/ai-agents-api): Function calling, tool use, MCP-compatible.
- [OCR API](https://www.runcrate.ai/ocr-api): Document text extraction.
- [Face detection API](https://www.runcrate.ai/face-detection-api): Vision-language face detection.
- [Background removal API](https://www.runcrate.ai/remove-background-api): AI-powered image isolation.
- [ML infrastructure](https://www.runcrate.ai/ml-infrastructure): Platform overview for AI teams.
## Solutions
- [Solutions index](https://www.runcrate.ai/solutions): Overview of every Runcrate use case.
- [Inference](https://www.runcrate.ai/solutions/inference): Production LLM serving, vLLM, TGI.
- [Fine-tuning](https://www.runcrate.ai/solutions/fine-tuning): LoRA, QLoRA, full fine-tuning on bare-metal GPUs.
- [Model training](https://www.runcrate.ai/solutions/model-training): Multi-node distributed training (DeepSpeed, FSDP, Megatron-LM).
- [RAG pipelines](https://www.runcrate.ai/solutions/rag): Embeddings + vector search + generation on one platform.
- [AI agents](https://www.runcrate.ai/solutions/ai-agents): GPUs your agent runs jobs on from Claude Code, Codex or Cursor over MCP. Each job's machine stops when its command ends; every job has a max runtime; a monthly budget caps the environment the agent is pinned to at sign-in.
- [Computer vision](https://www.runcrate.ai/solutions/computer-vision): Vision models for OCR, document analysis, video understanding.
- [Image and video generation](https://www.runcrate.ai/solutions/image-video): FLUX, Imagen, Veo 3, Kling.
- [Speech and audio](https://www.runcrate.ai/solutions/speech-audio): TTS (Qwen3 TTS, Kokoro, Orpheus) and STT (Whisper, Voxtral).
- [LLM API solutions](https://www.runcrate.ai/solutions/nlp-llms): DeepSeek, Llama, Qwen, gpt-oss via unified API.
- [Data processing](https://www.runcrate.ai/solutions/data-processing): RAPIDS, cuDF, Dask-CUDA on GPU clusters.
- [AI research](https://www.runcrate.ai/solutions/research): Jupyter, VS Code, per-minute billing for ablation studies.
- [Edge & embedded AI](https://www.runcrate.ai/solutions/edge-embedded): Train on cloud, deploy to Jetson / mobile / IoT.
## Documentation
- [Getting started](https://www.runcrate.ai/docs/g