Cerebras review: the fastest vendor-reported speeds of any AI inference provider
Answer in brief: Cerebras ranks #5 among the best AI inference providers. On the models it serves, it has the highest vendor-reported speeds we found: about 3,000 tokens/sec for gpt-oss-120b and about 1,850 tokens/sec for Qwen 3.8 27B, on its own wafer-scale systems. The self-serve Developer tier is pay-as-you-go ($0.35 / $0.75 per 1M tokens for gpt-oss-120b), but Cerebras labels it not intended for production. Production capacity is sold through Enterprise and Dedicated Inference.
Independent research, checked 2026-11-09 against Cerebras' pricing page and inference docs. Speeds are Cerebras' own figures; we did not test the product hands-on.
Wafer-scale speed for real-time AI
Cerebras runs inference on its wafer-scale systems rather than on conventional GPUs. For latency-critical work, such as coding agents waiting on long completions or reasoning models thinking in real time, its listed throughput is several times what most GPU clouds advertise. Cerebras also offers access through partners: AWS Marketplace, OpenRouter, Hugging Face and Vercel.
Who Cerebras fits best
Cerebras fits teams building latency-critical coding agents and real-time reasoning products on gpt-oss-120b or Qwen 3.8 27B. It also suits enterprises ready to sign a Dedicated agreement for other model families, custom weights or fine-tuning.
Why teams pick Cerebras
- Headline speed: about 3,000 tokens/sec on gpt-oss-120b and about 1,850 on Qwen 3.8 27B, as listed by Cerebras.
- Model fidelity: Cerebras says all Shared Inference models are original, unpruned versions. It uses only selective weight-only quantization in storage, and keeps activations, attention and the KV cache in full precision.
- A clear path up: Enterprise adds production-ready capacity, best priority, higher rate limits, enhanced latency options and a dedicated account team. Custom weights and fine-tuning are available by agreement.
- Many on-ramps: buy through AWS Marketplace, or reach it through OpenRouter, Hugging Face and Vercel.
- A free start: add a payment method to get a one-time $5 promotional credit.
The honest drawbacks
- Narrow shared catalog: only two models (gpt-oss-120b and Qwen 3.8 27B).
- Premium token pricing: gpt-oss-120b at $0.35 / $0.75, against $0.15 / $0.60 on several GPU clouds.
- Developer tier is not for production, according to Cerebras' own pricing page.
- The $5 credit expires 30 days after activation, and free-trial context is smaller (65k versus 131k paid for gpt-oss-120b).
Pricing at a glance
| Model | Speed (vendor) | Input / output per 1M | Context (free / paid) |
|---|---|---|---|
| gpt-oss-120b | ~3,000 tokens/sec | $0.35 / $0.75 | 65k / 131k |
| Qwen 3.8 27B | ~1,850 tokens/sec | $0.99 / $1.49 | 64k / 128k |
| Enterprise / Dedicated | n/a | Contact for a quote | n/a |
Best Cerebras alternatives
- Groq: fast inference with lower self-serve prices ($0.15 / $0.60 on gpt-oss-120b).
- Fireworks AI: Fast and Ultra-fast modes across a far larger catalog.
- SambaNova: custom-silicon inference with on-premises options.
Cerebras FAQ
Can I use Cerebras in production?
Yes, through Enterprise or Dedicated Inference. Cerebras says its self-serve Developer tier is for development, evaluation and experimentation, not production.
Does Cerebras prune or degrade models?
Cerebras says Shared Inference models are original and unpruned. It uses weight-only quantization in storage, with sensitive layers at full precision.
Which models does Cerebras serve?
Shared Inference serves gpt-oss-120b and Qwen 3.8 27B. Many more model families are available through Dedicated Inference.
Sources
- Cerebras inference pricing
- Cerebras model catalog
- Cerebras Dedicated Inference
← Back to the ranked list of the best AI inference providers
Named on this page as a provider? Join the network for a linked profile, Visit buttons and featured placement.