Model providersCerebras Systems

Cerebras

Decision matrix

Cerebras provider decision matrix
Overview

Cerebras runs a small set of open-weight models (currently GPT-OSS 120B and Qwen 3.8 27B on the self-serve tier) at high token-per-second speeds, with $5 of free credit to start and no ongoing free tier after that. Pick it when raw response speed on a supported model matters; pick a broader router when you need model variety.

Other good fits
  • Trying very fast inference on GPT-OSS 120B or Qwen 3.8 27B without buying dedicated hardware.
Pros
Read pros

Cerebras offers hosted inference on a short list of open-weight models through an API and a browser playground, running on its own accelerator hardware rather than general-purpose GPUs. The self-serve Developer tier currently covers GPT-OSS 120B and Qwen 3.8 27B; broader model access and production-grade capacity sit behind the Enterprise tier.

Pick it when…

  • You want to test whether high inference speed on a supported open-weight model actually helps your application feel more responsive.
  • One of the two Developer-tier models (GPT-OSS 120B or Qwen 3.8 27B) is a good fit for the task.
  • You'd rather not manage GPU hardware yourself just to try a fast open-weight model.
Cons
Read cons

A fast response to a short title request says nothing about quality on a long reasoning task; test with something closer to your real workload before committing. The Developer tier's model list is short by design, so if your task needs a specific model Cerebras doesn't host, you'll need a different provider regardless of speed.

Pick something else when…

  • You need a wide choice of models from one account. OpenRouter or Groq give you more model variety, at different speed and price trade-offs.
  • You need an ongoing free tier, not a one-time trial credit. Compare against providers that offer a small recurring free allowance.
  • EU-based infrastructure matters more than raw token speed. EUrouter is built around that instead.

Avoid if

  • You need a specific model Cerebras doesn't host; its self-serve catalogue is short by design.
Connection

Send a standard chat-completions request to https://api.cerebras.ai/v1/chat/completions with your API key as a bearer token and a supported model name (for example qwen-3.8-27b). Set max_completion_tokens to bound the response length; it limits generation length, not correctness.

Billing
Read billing

Prices checked 23 September 2026 (pricing (opens in a new tab)):

ModelSpeedInputOutput
GPT-OSS 120B~3,000 tokens/s$0.35 / M tokens$0.75 / M tokens
Qwen 3.8 27B~1,850 tokens/s$0.99 / M tokens$1.49 / M tokens

The Developer tier is self-serve, pay-as-you-go after a one-time $5 trial credit granted with a verified payment method; there's no recurring free allowance beyond that. The Enterprise tier adds the full model catalogue, production capacity and priority rate limits, but needs a sales conversation and a custom quote. Organisation-wide rate limits apply across your whole account, so another application on the same account can affect your throughput.

Data handling

Requests are processed on Cerebras's hosted infrastructure; there's no local or on-premises option on the self-serve tier. Check current terms before sending anything sensitive, since data-handling commitments differ between the Developer and Enterprise tiers.

First request

Documentation-based suggestion — not run here.

Send one short, limited request: ask for an eight-word title for a made-up event description. Check the response's choices[0].message.content against your input and note the reported token usage, not just the wall-clock speed.

Done when the title is accurate to the (fictional) event and you understand what the request used against your $5 trial credit.

Links & evidence
Documentation-based · not hands-on · Sep 23, 2026

Rests on Cerebras's quick start, pricing page and rate-limit documentation, read on 2026-09-23. No API key was created and no request was sent for this review.

  • Cerebras pricing · Sep 23, 2026 — Developer tier: GPT-OSS 120B $0.35/$0.75 per M tokens, Qwen 3.8 27B $0.99/$1.49 per M tokens, $5 trial credit.
  • Cerebras quick start · Sep 23, 2026 — API endpoint, example model, request/response structure.
  • Rate limits and trial terms · Sep 23, 2026 — One-time $5 trial credit after a verified payment method, no recurring free tier, organisation-level limits.
Further review notes & alternatives
Catalogue facts

Run open-weight models through an OpenAI-compatible inference API. Choose it when fast responses support interactive loops or batch jobs. Its practical boundary is open-weight model selection and smaller free-tier context.

Cerebras Systems · free-tier · cloud

  • fast inference
  • open-weight models
  • wafer-scale chip
  • api compatibility
  • generous free tier
  • real-time loops
  • automate
  • talk-think
  • write-code
  • draft-docs

Browse all tools