DeepInfra
Listed at https://deepinfra.com
💰 Token Pricing
| Type | Price | Note |
|---|---|---|
| Input | Llama 3.3 70B: $0.49/M, DeepSeek-V3: $0.99/M, DeepSeek-R1: $1.09/M | per million tokens |
| Output | Llama 3.3 70B: $0.73/M, DeepSeek-V3: $0.99/M, DeepSeek-R1: $1.09/M | per million tokens |
🤖 Supported Models (50)
✨ Pros
- ✓Highly competitive pricing (Llama 3.3 70B $0.49/M)
- ✓Rich model selection (50+ open-source)
- ✓Single API key for all models
- ✓Good reputation for API stability
- ✓Most models support Function Calling
- ✓OpenAI-compatible API, zero migration
⚠️ Cons
- ×China requires proxy
- ×No multimodal support
- ×Limited free credits
- ×Not a tier-1 brand, trust still building
- ×No enterprise SLA
🎯 Best For
Cost-sensitive open-source inference; multi-model development/testing; open-source Function Calling
💰 Pricing & Plans
| Model | Input ($/M) | Output ($/M) | Best For |
|---|---|---|---|
| Llama 3.3 70B Instruct | $0.49 | $0.73 | General open-source chat, low-cost reasoning |
| Llama 3.1 405B Instruct | $1.79 | $1.79 | Frontier-grade open-source inference at OpenAI-level |
| DeepSeek-V3 | $0.99 | $0.99 | Symmetric MoE — code + reasoning + RAG at flat pricing |
| DeepSeek-R1 | $1.09 | $1.09 | Reasoning-trace generation, chain-of-thought |
| Qwen2.5 72B Instruct | $0.49 | $0.49 | Bilingual EN/ZH chat, 128K context |
| Qwen2.5-Coder 32B Instruct | $0.18 | $0.18 | Cheapest coding model on the catalog |
| Mistral Large (2407) | $1.99 | $5.99 | European-hosted frontier Mistral inference |
| Mixtral 8x22B | $0.65 | $0.65 | MoE fallback when Llama is unavailable |
| All models — Free trial | Daily refresh, ~1k tokens | — | Sandbox testing without signup cost |
🔧 API & Developer Experience
- •API Style: OpenAI-compatible /chat/completions endpoint — drop-in replacement for OpenAI SDK clients. No separate DeepInfra SDK required.
- •Base URL: https://api.deepinfra.com/v1/openai — single global endpoint, no region pinning.
- •Model Selection: 50+ open-source models under one API key — switch via the `model` parameter without account changes or new credentials.
- •Streaming: SSE token streaming on every chat model; embedding models return batch arrays; image models stream progress over HTTP chunked.
- •Function Calling: Supported on Llama 3.x, Qwen2.5, Mistral, DeepSeek-V3 — JSON-schema tool definitions; not supported on legacy Mixtral or older WizardLM checkpoints.
- •Embeddings: BGE / E5 / sentence-transformers families via /v1/embeddings — same OpenAI-compatible shape; vector dimension varies per model.
- •Async Inference API: Long-context / batch jobs accepted via the async endpoint; results pollable for up to 24 hours with a request_id.
🏗️ Day-One Model Hosting
DeepInfra's core product is day-one hosting — making newly-released open-source weights available via API within hours of public release, often before tier-1 hyperscalers. The 2026 catalog demonstrates this pattern: Llama 3.3 70B, DeepSeek-V3, DeepSeek-R1, and Qwen2.5 family all shipped on DeepInfra within 1-7 days of upstream release. For developers who want to test a new model before committing to a Bedrock / Vertex / Azure contract, DeepInfra is the lowest-friction path: no enterprise sales call, no monthly minimum, no region restrictions. Pricing follows a flat per-million-token structure regardless of context length consumed — there is no premium tier for >100K context, no batch discount cliff, no provisioned-throughput contract. The trade-off is that DeepInfra runs on commodity GPU pools (mostly H100 + H200) and does not publish a formal SLA; uptime is best-effort, typically 99.5-99.9%, and burst capacity can be constrained when a popular model launches. For production workloads, the typical pattern is to prototype on DeepInfra, then port the same OpenAI-compatible request to Bedrock / Vertex for the SLA tier.
🌐 Regional Availability & Latency
DeepInfra's API (api.deepinfra.com) is not directly reachable from mainland China — the inference endpoints and the async polling endpoint both require stable proxy or VPN. Latency from non-CN regions is consistent: typical time-to-first-token is 250-400 ms for Llama 70B-class models on H100 hardware, with throughput of ~80 tokens/sec for streaming completions. There is no region pinning or multi-region routing — DeepInfra runs a single global pool, so latency is a function of the caller's path to the nearest US/EU edge rather than a regional cluster. For developers in Beijing / Shanghai / Shenzhen, the typical workaround is one of: (1) route through a US-region VPS that already proxies OpenAI / Anthropic traffic, (2) use an aggregator (OpenRouter, Portkey, Cloudflare AI Gateway) that ships a CN edge node and signs the upstream contract on your behalf, or (3) self-host the same open-source weights on Aliyun PAI / Tencent TI-ONE if data-residency in China is the primary constraint. DeepInfra does not offer an enterprise data-residency tier.