Cerebras
Fast LLM inference on wafer-scale hardware
Last reviewed Sep 5, 2026
Cerebras Systems builds wafer-scale AI processors and runs Cerebras Inference, a token-billed LLM API served on its own CS-3 systems rather than GPUs. The public endpoints expose a small catalog of open-weight models through an OpenAI-compatible API, with private dedicated endpoints available for reserved capacity and additional model families.
LLM API Pricing
Pay-per-token pricing. Prices shown per 1M tokens.
Prices last updated: October 20, 2026
| Model | Input/1M | Output/1M |
|---|---|---|
| $0.350 | $0.750 | |
| $0.990 | $1.49 |
Cerebras pricing by model
Input, output, and batch rates, plus alternatives, for one model at a time.
Pros & Cons
Advantages
- Published throughput in the thousands of tokens per second on public models
- Drop-in OpenAI-compatible API with first-party Python and TypeScript SDKs
- Serves original, unpruned open-weight models with a stated no-modification policy
- Free trial credits on signup, then self-serve pay-as-you-go
- Per-model documentation lists context, output limits, and rate limits per tier
Limitations
- Very small public model catalog compared with GPU-based inference providers
- Free-tier context windows are roughly half the paid-tier limits
- Preview models can be withdrawn and are not intended for production use
- Larger model families require a dedicated endpoint quoted through sales
- Inference only: no GPU rental, training, or self-serve fine-tuning
Key Features
Wafer-Scale Inference
Models run on the Cerebras Wafer-Scale Engine, which keeps model weights in on-chip SRAM to reach published speeds in the thousands of tokens per second
OpenAI-Compatible API
Chat Completions and Completions endpoints at api.cerebras.ai/v1, usable from the OpenAI SDKs or the Cerebras Python and TypeScript SDKs
Open-Weight Model Catalog
Public endpoints serve unmodified open-weight models such as OpenAI GPT OSS and Qwen, with a documented policy of no pruning on hosted models
Reasoning and Tool Calling
Configurable reasoning effort, streaming, structured outputs, parallel tool calling, and prompt caching across the catalog
Image Inputs
Vision-capable models accept PNG and JPEG images alongside text
Dedicated Endpoints
Private, provisioned endpoints on reserved capacity with fine-tuning, weight management, and additional model families, arranged through sales
Cloud Console and Playground
Web console for API keys, usage and cost monitoring by model, and an interactive playground
Pricing Options
| Option | Details |
|---|---|
| Free Trial | Time-limited credits granted on signup, with lower rate limits and context windows than paid tiers |
| Developer (Pay as You Go) | Prepaid credits billed per million input and output tokens at published per-model rates |
| Enterprise and Dedicated Endpoints | Reserved capacity, higher rate limits, and additional models on custom terms through sales |
Availability & Support
Regions
Cerebras-operated data centers in North America, with expansion into Europe; no region selection on the public API
Support
Documentation and API reference, model-specific guides, Cloud Console usage monitoring, Discord community, and enterprise support for dedicated customers
Getting Started
- 1
Create an account
Sign up at cloud.cerebras.ai and add a payment method to activate the free trial
- 2
Get an API key
Generate a key from the Cloud Console
- 3
Install an SDK
pip install cerebras_cloud_sdk or npm install @cerebras/cerebras_cloud_sdk, or use the OpenAI SDK pointed at api.cerebras.ai/v1
- 4
Make a request
Call chat completions with a model ID from the catalog, such as gpt-oss-120b
Frequently Asked Questions
Which models does Cerebras serve?
Check the pricing table above for the models Cerebras currently serves and their per-token rates.
How do I get started with Cerebras?
Create an account, Get an API key, Install an SDK, Make a request
What are Cerebras's main advantages?
Cerebras's main advantages include: Published throughput in the thousands of tokens per second on public models, Drop-in OpenAI-compatible API with first-party Python and TypeScript SDKs, Serves original, unpruned open-weight models with a stated no-modification policy, Free trial credits on signup, then self-serve pay-as-you-go, Per-model documentation lists context, output limits, and rate limits per tier.
What are Cerebras's limitations?
Cerebras's main limitations include: Very small public model catalog compared with GPU-based inference providers, Free-tier context windows are roughly half the paid-tier limits, Preview models can be withdrawn and are not intended for production use, Larger model families require a dedicated endpoint quoted through sales, Inference only: no GPU rental, training, or self-serve fine-tuning.