Cerebras is an inference cloud serving open-weight language models through an OpenAI-compatible API, billed per token, for developers building fast agents.
Cerebras returns language model output fast enough for a coding agent or a voice assistant to keep pace with the person using it. The service is sold as Cerebras Inference and is owned by Cerebras Systems Inc. of Sunnyvale, California. It runs open-weight models on the company's own wafer-scale chips. Requests follow the OpenAI request format, so code written for OpenAI's SDKs can move across with few changes. Developers sign up in the Cloud console, create API keys and try prompts in the Playground before writing code. The API offers Chat Completions and Completions endpoints, with GPT OSS 120B and Qwen 3.8 27B on shared inference. Tool calling supports strict mode and parallel calls, and Structured Outputs hold a reply to a supplied JSON Schema. A reasoning_effort setting trades speed for deeper thinking on Qwen, which also reads PNG and JPEG images. Dedicated Inference hosts further models and a customer's own weights. The service suits developers and AI engineering teams whose product slows down while a model generates text. Individual builders start with a trial credit and then buy pay-as-you-go credits, billed per token, on a Developer tier meant for development and evaluation. Companies that put an application into production sign an Enterprise agreement by quote, which brings priority capacity, fine-tuning services and an account team. Prompts and responses are not retained by default, with no separate contract needed. Most of the work is agentic: coding assistants in Cline, research agents that chain many calls, and voice agents built with LiveKit that must answer mid-conversation. Teams writing their pipelines in LangChain or LlamaIndex can select Cerebras as the model provider, and LiteLLM or OpenRouter route traffic to it beside other providers. Buyers who already commit cloud spend through AWS Marketplace can pay for capacity there. Cerebras fits wherever waiting on generated tokens is the slowest step in the loop.
Pricing
View plans-
Free trial credit
Free
- One-time $5 promotional credit after adding a payment method
- Credit expires 30 days after activation; no automatic charge or enrollment in paid usage
- 5 requests per minute and 1M tokens per day on each model
- Context 65k on GPT OSS 120B and 64k on Qwen 3.8 27B, about half the paid context
- Up to 2 images per request on Qwen 3.8 27B
- API and Playground pause when the credit expires or runs out unless PayGo credits are bought
- Community Discord support
-
Developer
Custom pricing
- Pay as you go per million tokens, no flat fee and no billing toggle
- GPT OSS 120B: $0.35 per million input tokens, $0.75 per million output tokens
- Qwen 3.8 27B: $0.99 per million input tokens, $1.49 per million output tokens
- For development, evaluation and experimentation; not intended for production use
- GPT OSS 120B at 1K requests and 3M tokens per minute
- Qwen 3.8 27B at 300 requests and 750K tokens per minute; no hourly or daily caps
- Context 131k on GPT OSS 120B and 128k on Qwen 3.8 27B
- Up to 10 images per request on Qwen 3.8 27B
- Cached input tokens billed at the standard input rate
- Only these two models on shared inference; other models need Dedicated Inference
- Credits bought in advance; auto-recharge is off by default
- Community Discord support
-
Enterprise
Custom pricing
- Quote only, contact sales
- All supported models with production-ready capacity
- Best priority, higher rate limits provisioned per organization, enhanced latency options
- Custom model weights and fine-tuning and training services, by agreement
- Dedicated Inference deployments
- Dedicated account team
Features
- Included: Published context window size
- Included: Image input support
- Not included: Audio input support
- Included: Function and tool calling
- Included: JSON schema output mode
- Included: Extended reasoning mode
- Not included: Prompt caching discount
- Not included: Batch inference API
- Included: Fine-tuning offered
- Included: Open weights release
- Included: Zero data retention option
- Not included: Regional data residency
- Not included: Published safety evaluations
- Included: Free tier or trial credits
- Included: Enterprise agreement available
- Not included: Published deprecation policy
- Included: Published benchmark results
- Not included: Embeddings endpoint
Additional Features
- OpenAI-compatible Chat Completions API
- Completions endpoint
- Self-serve Cloud console
- Playground for trying prompts
- GPT OSS 120B on shared inference
- Qwen 3.8 27B on shared inference
- Tool calling with strict mode
- Parallel tool calls on Qwen 3.8 27B
Best for
- developers building coding agents that need fast token output
- teams prototyping voice agents where response delay matters
- engineers moving an OpenAI SDK app onto open-weight models
- developers evaluating open models before a production contract
- enterprises that need dedicated capacity for custom or tuned weights
- teams that want prompts and outputs left unstored by default
- buyers who already commit cloud spend through AWS Marketplace
Use cases
- running an agentic coding assistant in Cline
- powering a real-time voice agent built on LiveKit
- research agents that chain many model calls in a row
- pointing an existing OpenAI SDK app at a new base URL
- reaching GPT OSS 120B through OpenRouter or LiteLLM
- returning schema-valid JSON to downstream code
- answering questions about screenshots and photos
- building LangChain or LlamaIndex pipelines on open models
- streaming answers in a Vercel AI SDK web app
- serving fine-tuned weights on Dedicated Inference
User Reviews
Compare with Similar Software
Click to compare side by side with Cerebras
