DigitalOcean Gradient
Listed at https://gradient.digitalocean.com
💰 Token Pricing
| Type | Price | Note |
|---|---|---|
| Input | Llama 3.3 70B: $0.50/M, DeepSeek-R1: $0.80/M | per million tokens |
| Output | Llama 3.3 70B: $0.50/M, DeepSeek-R1: $0.80/M | per million tokens |
🤖 Supported Models (15)
✨ Pros
- ✓DigitalOcean brand reliability, high developer trust
- ✓$100 free credits for new users
- ✓OpenAI-compatible API
- ✓Transparent and competitive pricing
- ✓NVIDIA H100 GPU infrastructure
- ✓Joined OpenRouter provider network May 2026
⚠️ Cons
- ×China requires proxy
- ×Limited model selection (15 open-source)
- ×No multimodal input
- ×New on OpenRouter, reputation building
- ×No proprietary models, relies on open-source community
🎯 Best For
Existing DigitalOcean customers; developers needing $100 free credits for experimentation; open-source model deployment
💰 Pricing & Plans
| Model | Input ($/M) | Output ($/M) | Notes |
|---|---|---|---|
| Llama 3.3 70B | $0.50 | $0.50 | Balanced flagship via Gradient + OpenRouter (digitalocean/llama-3.3-70b-instruct) |
| Llama 3.1 405B | $0.90 | $0.90 | Largest open-weights generalist on Gradient |
| DeepSeek-R1 | $0.80 | $0.80 | Reasoning model; also routed via OpenRouter |
| Qwen 2.5 72B | — | — | Via Gradient serverless; billed per token |
| DeepSeek V4 Flash 0731 | — | — | Added Aug 6, 2026 to DeepSeek-family lineup |
| DeepSeek V4 Pro 0813 | — | — | Added Aug 14, 2026; agentic + 1M context |
🔧 API & Developer Experience
- •API Surface: Gradient exposes an OpenAI-compatible chat completions endpoint (https://inference.do-ai.run/v1/). The standard OpenAI SDK works by swapping base_url and api_key. On OpenRouter the models are reachable under the digitalocean/ model prefix (e.g. digitalocean/llama-3.3-70b-instruct) with no separate DigitalOcean account.
- •Signup & Credits: Create an account at cloud.digitalocean.com, add a payment method, and new users receive $200 in credits valid 60 days. No sales call; the platform is fully self-serve.
- •Serverless vs GPU Droplets: The Gradient AI Platform handles autoscaling, batching and routing for hosted models (pay per token), while GPU Droplets give raw H100/H200/L40S/A100 boxes by the hour for you to run vLLM/TGI/SGLang or custom stacks.
- •Function Calling & Agents: The chat completions endpoint supports function calling / tool use. The Agent platform adds hosted tool bindings (web search, calculator, code execution) plus a managed Knowledge Base vector store that chunks and embeds your PDFs/docs.
- •API Key Management: Generate scoped API tokens from the DO control panel; tokens work across GPU Droplets, serverless inference, Spaces and the rest of the DigitalOcean ecosystem.
🤖 Open-Weights Catalog & OpenRouter Hookup
DigitalOcean Gradient is two products glued together: GPU Droplets (hourly H100/H200/L40S/A100 virtual servers) and the Gradient AI Platform, a serverless inference and agent layer relaunched in late 2025. The platform hosts a curated open-weights catalog (Llama 3.3 70B, Llama 3.1 405B, DeepSeek-R1, Qwen 2.5 72B, Mixtral, Mistral Large 2, Gemma 2, Phi-3) with no proprietary models - every model is community open weights. The June 3, 2026 OpenRouter integration made these models reachable to every OpenRouter user through the digitalocean/ model prefix without a DigitalOcean account, effectively turning DigitalOcean into an anonymous open-model inference back end for the broader API ecosystem. In August 2026 the catalog extended to the DeepSeek V4 family: V4 Flash 0731 (Aug 6) and V4 Pro 0813 (Aug 14) joined serverless inference, bringing 1M-context and stronger agentic tool-calling to the platform. The Knowledge Base and Agent features sit on top of the serverless layer, handling the boring RAG plumbing in the same control panel as your droplets.
🌐 Regional Availability & Latency
DigitalOcean Gradient inference runs from DigitalOcean's global network across North America and Europe data centers, so serverless requests land close to your droplets and app servers. The OpenAI-compatible endpoint keeps client-side integration latency low. For mainland China there is no regional endpoint and no official China access program: the dashboard can be blocked and API calls need routing through a stable overseas proxy or an aggregator. OpenRouter's China-direct edge is a practical workaround - routing to digitalocean/llama-3.3-70b-instruct through OpenRouter lets mainland developers reach the same models with a better-connected path. For teams that control latency directly, GPU Droplets can be pinned to a chosen region so vLLM/TGI serving and the rest of the stack are co-located, giving predictable p95 latency for production workloads.