What it is
Cerebras builds very large AI chips, each made from a whole silicon wafer, and rents them out through an inference API. It is usually the fastest option for the models it serves: its docs list gpt-oss-120b at about 3,000 tokens per second and Qwen 3.8 27B at about 1,850.
The shared API lists only those two models. Its pricing page puts gpt-oss-120b at $0.35 per million input tokens and $0.75 per million output, and Qwen 3.8 27B at $0.99 and $1.49. New accounts get a one-time $5 credit after adding a card, valid for 30 days; when it runs out, access stops until you buy pay-as-you-go credits. More models, reserved capacity and SLAs come through Dedicated Inference by agreement. Cerebras went public on Nasdaq in May 2026 and has a large compute deal with OpenAI.
Our review
Cerebras is the speed specialist among inference APIs, and both the speed and the narrowness are real. Its docs list gpt-oss-120b at about 3,000 tokens per second, so a long answer arrives almost at once, and for coding agents that chain dozens of calls the saved seconds add up to minutes per task.
The shared API offers two models, gpt-oss-120b and Qwen 3.8 27B; anything else means Dedicated Inference and an enterprise agreement. Cerebras charges $0.35 per million input tokens and $0.75 output for gpt-oss-120b, while OpenRouter lists Groq at $0.15 and $0.60 for the same model and several smaller hosts lower still. A month of 2 billion input and 500 million output tokens costs about $1,075 on Cerebras against $600 on Groq.
The trial tier is for evaluation only: 5 requests a minute and 1 million tokens a day. Pay-as-you-go lifts gpt-oss-120b to 1,000 requests and 1 million uncached tokens a minute with no daily cap, but Qwen 3.8 27B stays at 300 requests a minute. When the $5 credit expires after 30 days, API access stops rather than rolling into paid billing, which is a sensible default for a trial.
On data, Cerebras says it does not retain prompts, requests or outputs from the inference API, keeping only account and usage data, though prompt-cache entries live for at least 5 minutes and up to an hour. A SOC 2 Type 2 report and DPA are available on request through its trust centre, and the DPA covers business accounts, not individual sign-ups. The fixed-price Cerebras Code plans at $50 and $200 a month are both marked sold out on Cerebras's own Code page, which leaves the per-token API as the dependable way in.
Where it earns its place
A developer-tools startup running a coding agent with dozens of sequential model calls per task
Cut wall-clock time per agent task
gpt-oss-120b at about 3,000 tokens a second shortens every step, and 1,000 requests a minute on pay-as-you-go is rarely the bottleneck. For an input-heavy mix expect to pay about 1.8 times Groq's rate.
A team building real-time voice agents for phone support
Keep spoken replies under a second
Fast generation leaves more of the latency budget for speech recognition and synthesis. Qwen 3.8 27B, at about 1,850 tokens a second and $0.99/$1.49 per million, accepts images too, but its 300-requests-a-minute cap matters at contact-centre volume.
A head of engineering weighing inference vendors
Decide whether speed justifies a premium and a narrow catalogue
Routing only latency-critical calls to Cerebras through OpenRouter avoids a second contract and keeps the rest of the traffic on cheaper hosts; the case weakens if the model you need is only on Dedicated Inference.
Before you commit
- Setup
- Sign up, add a card for the $5 credit and call the OpenAI-compatible endpoint; the trial tier is labelled for evaluation, not production.
- Integrations
- OpenAI-compatible API; also sold through OpenRouter, Hugging Face, AWS Marketplace and Vercel.
- Data and privacy
- Prompts, requests and outputs not retained per Cerebras; prompt cache kept at least 5 minutes and up to an hour; SOC 2 Type 2 report and DPA on request from the trust centre.
- Cost at scale
- 2 billion input and 500 million output tokens a month on gpt-oss-120b cost about $1,075; the same mix at Groq's OpenRouter price is about $600.
- Limits
- Context 131k on gpt-oss-120b and 128k on Qwen 3.8 27B for paid accounts (65k and 64k on the trial); Qwen capped at 300 requests a minute on pay-as-you-go.
What's good
- About 3,000 tokens/s on gpt-oss-120b, per its docs
- Prompts and outputs not retained, per Cerebras
- 1,000 requests a minute and no daily cap on pay-as-you-go
- Also sold through OpenRouter, Hugging Face, AWS Marketplace and Vercel
What's not
- Only two models on the shared API
- gpt-oss-120b input costs more than twice Groq's
- Trial capped at 5 requests a minute
- Cerebras Code plans listed as sold out
Best for
Developers building latency-sensitive agents or coding tools that can work with gpt-oss or Qwen models.
Alternatives
Cerebras: common questions
- Is Cerebras worth it?
- Cerebras sells speed by the token, and at more than twice Groq's input price for gpt-oss it only pays off when latency is the product. We rate it 4.0 out of 5.
- How much does Cerebras cost?
- $5 one-time trial credit (30 days), then pay-as-you-go. gpt-oss-120b $0.35 in / $0.75 out per 1M tokens; Qwen 3.8 27B $0.99 / $1.49. Dedicated Inference by agreement. Cerebras Code subscriptions ($50 and $200/month) listed as sold out.
- What are the downsides of Cerebras?
- Only two models on the shared API; gpt-oss-120b input costs more than twice Groq's; trial capped at 5 requests a minute; cerebras Code plans listed as sold out.
- Who is Cerebras best for?
- Developers building latency-sensitive agents or coding tools that can work with gpt-oss or Qwen models.
- How much does Cerebras cost for a team or in production?
- 2 billion input and 500 million output tokens a month on gpt-oss-120b cost about $1,075; the same mix at Groq's OpenRouter price is about $600.
- Is Cerebras safe for company data?
- Prompts, requests and outputs not retained per Cerebras; prompt cache kept at least 5 minutes and up to an hour; SOC 2 Type 2 report and DPA on request from the trust centre.
- What are the best Cerebras alternatives?
- The closest alternatives we have reviewed are Groq, Fireworks AI, Together AI and OpenRouter.
Answers checked Oct 23, 2026.
Make Cerebras? Show this rating on your site
