Cerebras
Listed at https://inference.cerebras.ai
💰 Token Pricing
| Type | Price | Note |
|---|---|---|
| Input | $0.10/M to $0.60/M tokens（所有模型统一 $0.60/M input） | per million tokens |
| Output | $0.10/M to $0.60/M tokens（统一 $0.60/M output） | per million tokens |
🤖 Supported Models (7)
✨ Pros
- ✓WSE-3 wafer-scale chip, 2000+ tok/s inference
- ✓Llama 3.3 70B reaches 2,000+ tok/s, industry fastest
- ✓Zero cold start
- ✓OpenAI-compatible API
- ✓Reasonable pricing ($0.60/M tokens flat)
- ✓GPU alternative with unique inference chip design
⚠️ Cons
- ×China requires proxy
- ×Limited model selection (only 6)
- ×No Function Calling support
- ×Speed dependent on Llama/Qwen base model performance
- ×Non-general-purpose inference chip, limited third-party support
- ×No multimodal models
🎯 Best For
Applications demanding extreme inference speed; high-throughput real-time scenarios; high-performance Llama inference
💰 Pricing & Plans
| Model / Tier | Input ($/M tokens) | Output ($/M tokens) | Speed | Notes |
|---|---|---|---|---|
| Free Trial | $0 (with $5 credits on signup) | $0 (with $5 credits on signup) | Full speed | All Cerebras-powered models; community Discord support; expires when credits run out |
| Developer (Pay-as-you-go) | From $10 self-serve top-up | — | 10x Free Trial rate limits; higher priority | Self-serve at cloud.cerebras.ai; everything in Free plus priority processing |
| Enterprise (Sales-led) | Custom contract | Custom contract | Reserved capacity; custom weights | Highest rate limits, dedicated queue priority, fine-tuning support, SLA |
| GPT-OSS 120B (openai/gpt-oss-120b) | $0.35 | $0.75 | ~3,000 tok/s | Public endpoint flagship; 65k ctx (free) / 131k ctx (paid); 32k free / 40k paid max output |
| Gemma 4 31B (gemma-4-31b) | $0.99 | $1.49 | ~1,850 tok/s | Multimodal (text + image input); 65k free / 131k paid ctx; supports parallel tool calling |
| Llama 3.3 70B (legacy / Dedicated Endpoints) | Contact Sales | Contact Sales | Dedicated capacity | Removed from public endpoints in 2026; reserved capacity only via Dedicated Endpoints |
🔧 API & Developer Experience
- •OpenAI-compatible Chat Completions: /v1/chat/completions endpoint mirrors OpenAI's API — drop-in for any OpenAI SDK by pointing the base URL at https://api.cerebras.ai/v1; standard parameters (temperature, top_p, frequency_penalty, presence_penalty, seed, logit_bias) supported.
- •Fastest inference in the industry: Measured at ~3,000 tokens/sec on GPT-OSS 120B and ~1,850 tokens/sec on Gemma 4 31B; first-token latency under 100 ms because the entire model lives on a single wafer-scale chip rather than split across GPUs.
- •Native Python & Node.js SDKs: cerebras.cloud.sdk (Python) and the Node.js client wrap the REST API; free tier gets 5 requests/min and 30k input tokens/min with a 1M daily cap; Developer tier raises that to 1,000 requests/min and 1M tokens/min.
- •Standard LLM capabilities: Reasoning (configurable reasoning_effort: low/medium/high), streaming responses, structured outputs (JSON schema), tool calling, prompt caching, and sampling controls — all available on both GPT-OSS 120B and Gemma 4 31B.
- •Multimodal image inputs: Gemma 4 31B accepts images in Chat Completions (text + image → text) for visual agentic workflows, document Q&A and screenshot-to-code; up to 10 images per request on Developer tier (2 per request on Free).
- •Dedicated Endpoints for reserved capacity: Beyond the public shared endpoints, Cerebras offers Dedicated Endpoints for Llama/Qwen legacy models and custom weights — guaranteed capacity, dedicated queue priority, fine-tuning support, and 24x7 SLA-backed support.
🧠 Wafer-Scale Inference & WSE-3
Cerebras' signature capability is the Wafer-Scale Engine-3 (WSE-3), a single silicon die the size of a dinner plate that hosts an entire large language model rather than splitting it across many GPUs. With 44 GB of on-chip SRAM and roughly 900,000 AI-optimized cores connected by a fabric that delivers tens of petabytes per second of bandwidth, the WSE-3 holds the world's largest models fully on-chip and eliminates inter-GPU communication overhead entirely. Measured inference speeds reach ~3,000 tokens/second on GPT-OSS 120B and ~1,850 tokens/second on Gemma 4 31B — roughly 20x faster than OpenAI or Anthropic endpoints for streaming workloads. Because the model never leaves the wafer, the first-token latency stays under 100 ms even on long prompts, which is what makes Cerebras uniquely well-suited to interactive coding assistants, voice agents and real-time document Q&A. The Dedicated Endpoints program extends the same wafer to custom weights, letting enterprises fine-tune or run private-model deployments on the same physical substrate.
🌐 China Access & Latency
api.cerebras.ai is hosted entirely in Cerebras' US data centers, and mainland China traffic is not directly reachable from Chinese ISPs without a stable proxy. The free-trial signup at cloud.cerebras.ai also loads slowly from Chinese networks; new developers commonly report 5-30 second connect times before the page resolves, and the dashboard / playground can intermittently time out. Once the API key is obtained via proxy, request latency from a China-based server is typically 200-500 ms round-trip depending on the proxy route — much slower than the sub-100 ms first-token time that Cerebras advertises from US endpoints. For China-resident production deployments the practical options are: run the workload on a Hong Kong / Singapore / Tokyo proxy with reserved capacity on Dedicated Endpoints; or use Cerebras via partner channels (OpenRouter, Hugging Face, AWS Marketplace, or Vercel) which provide mirrored endpoints with regional routing. China-based startups evaluating Cerebras for voice, coding, or agentic real-time workloads should budget the proxy latency into their p99 SLOs and design fallback paths to a domestic provider when the proxy underperforms.