Cerebras Inference
Listed at https://inference.cerebras.ai
💰 Token Pricing
| Type | Price | Note |
|---|---|---|
| Input | Gemma 4 31B $0.99/M, Z.ai GLM 4.7 $2.25/M, OpenAI GPT-OSS 120B $0.35/M, Llama 3.3 70B $0.60/M | per million tokens |
| Output | Gemma 4 31B $1.49/M, Z.ai GLM 4.7 $2.75/M, OpenAI GPT-OSS 120B $0.75/M, Llama 3.3 70B $0.60/M | per million tokens |
🤖 Supported Models (7)
✨ Pros
- ✓CS-4 wafer-scale chip delivers 30× faster batch-1 inference than merchant GPUs
- ✓Fastest hosted inference for the open-weights OpenAI GPT-OSS 120B (Apache-2.0)
- ✓Multimodal Gemma 4 31B — first vision-capable Cerebras-hosted model
- ✓Full OpenAI-compatible API surface: function calling + streaming + JSON mode
- ✓$5 free trial credits + $10 Developer top-up unlocks the full model catalog
⚠️ Cons
- ×US-only deployment; Asia-Pacific TTFT 600-900 ms is acceptable for batch workloads but noticeable in interactive chat
- ×No native CN access — mainland developers require a proxy or a domestic fallback provider
- ×Catalog skews toward open-weights (GPT-OSS / Llama / Gemma / GLM); closed frontier models like Claude / GPT / Gemini are absent
- ×No standing free tier — the $5 trial is one-time, not recurring
- ×OpenAI-style tool-use conventions; Anthropic-style function calling ergonomics require porting
🎯 Best For
Latency-sensitive interactive agents, in-IDE code completion, and streaming chat; teams needing managed inference for open-weights (GPT-OSS 120B / Llama 3.1 405B); applications that benefit from direct Gemma 4 multimodal input.
💰 Pricing & Plans
| Tier | Top-up | Rate Limits | Model Access |
|---|---|---|---|
| Free Trial | $5 signup credit | Lowest (free-tier) | All Cerebras-hosted models (Gemma 4 multimodal, GLM 4.7, GPT-OSS 120B, Llama 3.3/3.1) |
| Developer | $10 minimum top-up | 10× the free tier | Same catalog; higher priority processing |
| Enterprise | Custom (sales-led) | Dedicated queue; SLA-backed | Custom model weights + fine-tuning services; uptime guarantees |
🔧 API & Developer Experience
- •OpenAI-Compatible REST API: Drop-in `/v1/chat/completions` endpoint with OpenAI schema — existing OpenAI SDKs (Python, Node, Go) work by changing only the `base_url` to `https://inference.cerebras.ai/v1` and supplying a Cerebras API key, so migration requires zero code changes.
- •Streaming by default: Server-sent events (SSE) and JSON stream mode both supported on every chat model; first-token latency is sub-second for Gemma 4 31B and ~200 ms for smaller Llama variants on a warm connection.
- •Function Calling / JSON Mode: Native tool-use on GPT-OSS 120B and Z.ai GLM 4.7; `response_format={"type":"json_schema",...}` enforced JSON output on every text-capable model; parallel tool execution supported.
- •Multimodal image input: Gemma 4 31B accepts base64 image inputs (PNG, JPEG, WebP up to 20 MB) directly in chat-completion messages — first-party screenshots, document scans, and diagrams without separate OCR; multimodal routing is automatic.
- •Inference API key scoped to single project: API keys are scoped to a per-project ID and rotate without service interruption; per-key usage telemetry surfaces in the Cerebras dashboard under a per-model breakdown.
- •Python SDK + LlamaIndex integration: Official `cerebras-cloud-sdk` Python package mirrors the OpenAI client surface; LlamaIndex and LangChain integrations via `ChatCerebras` adapter (in langchain-cerebras) — no glue code required for RAG pipelines.
🎯 Wafer-Scale CS-4 Hardware & Frontier Model Catalog
Cerebras Inference is one of the few platforms that runs production traffic on wafer-scale CS-4 silicon rather than merchant GPUs. The wafer's 30× speed-up at batch-1 means the inference service delivers frontier-tier latency on Llama 3.1 405B and OpenAI gpt-oss-120b at rates that traditional GPU clouds cannot match. The 2026 Q2/Q3 catalog added multimodal Gemma 4 31B (the first vision-capable Cerebras model), Z.ai GLM 4.7 for coding/reasoning, and OpenAI gpt-oss-120b under Apache-2.0 — all priced at sub-$2.50 per million output tokens. For workloads where batch-1 latency matters — interactive agents, real-time autocomplete, code completion inside an IDE, streaming chat — Cerebras is the only hosted inference platform where a 405B-class model returns the first token in under a second.
🌐 Regional Availability & Latency
Cerebras Inference is hosted exclusively from `inference.cerebras.ai`, served from US-based CS-4 hardware. North America customers typically see 50-200 ms TTFT on small models and ~400 ms on 405B-class models. European customers reach the US origin over trans-Atlantic fiber with TTFT roughly 1.5× the North America baseline. Asia-Pacific customers (Tokyo / Singapore / Sydney) see 600-900 ms TTFT on the same workloads — acceptable for batch pipelines but noticeable in interactive chat. Mainland China is the explicit exception: Cerebras does not currently serve inference traffic from any in-China region and outbound traffic from CN is frequently throttled by the GFW, so mainland developers should expect either 1.5 s+ TTFT or outright connection failures — a domestic provider (Zhipu, Qwen, Doubao) is the recommended fallback for user-facing applications serving CN traffic.