Verdict and evidence boundary
Cerebras sells LLM inference running on its own wafer-scale hardware. Cerebras describes the WSE-3 in its CS-3 system as a single wafer-scale chip with 46,225 mm² of silicon, 900,000 cores, and 44 GB of on-chip SRAM, and that design is the basis for its speed claims.
This review is Documented. It summarizes Cerebras documentation, its pricing page, its published performance figures, and the independent Artificial Analysis leaderboard. aicoolies has not run its own latency or throughput tests, so every performance figure below is attributed to its source.
Hardware Architecture: The Wafer-Scale Engine (CS-3)
On GPU clusters, token-by-token generation is often limited by moving model weights and KV cache between external HBM memory and compute. Cerebras keeps weights and activations in on-chip SRAM on one wafer-sized chip, which removes most off-chip memory traffic. Cerebras's datasheet lists 21 PB/s of on-chip memory bandwidth for the WSE-3, and the company claims inference up to 30 times faster than GPU-based systems.
Published Speed Figures
Cerebras lists about 3,000 tokens per second for gpt-oss-120b and about 1,850 tokens per second for Qwen 3.8 27B on its pricing page and in its documentation (checked 28 Sep 2026). Independently, Artificial Analysis ranked Cerebras fastest of 18 providers for gpt-oss-120b (72-hour median, as of 28 Sep 2026). These figures change as models and hardware are updated, so readers should check both sources for current numbers. aicoolies did not measure time-to-first-token, throughput, or end-to-end latency.
Developer Ergonomics and OpenAI SDK Compatibility
Cerebras documents an OpenAI-compatible API. Teams using the OpenAI SDK point their client at `https://api.cerebras.ai/v1` and supply a Cerebras API key (`client = OpenAI(base_url='https://api.cerebras.ai/v1', api_key=...)`).
The documentation covers tool calling, structured outputs with JSON schema, streaming, and sampling parameters such as temperature (0 to 2) and top_p (0 to 1). One documented limitation is that `json_object` mode cannot be combined with streaming. This makes Cerebras usable from LangChain, LlamaIndex, Pydantic AI, and custom agent loops through their OpenAI-compatible clients. aicoolies has not measured schema adherence or tool-call reliability.
Pricing, Model Catalog, and Trade-offs
The Developer tier is pay-as-you-go and starts with a free $5 credit. Cerebras lists gpt-oss-120b at $0.35 per million input tokens and $0.75 per million output tokens, and Qwen 3.8 27B at $0.99 and $1.49; Enterprise pricing is by quote (checked 28 Sep 2026).
The main trade-off is catalog scope. Cerebras's own API serves open-weights models rather than proprietary models such as Claude. Context limits vary by model and plan: gpt-oss-120b is listed at 65K tokens on the free tier and 131K on paid tiers, and Qwen 3.8 27B at 64K and 128K. Teams that need very long contexts should check these limits before committing.
Final Verdict
Cerebras is worth evaluating first when generation speed is the main constraint and an open-weights model fits the task. Artificial Analysis ranked it fastest of 18 providers for gpt-oss-120b (72-hour median, as of 28 Sep 2026), and the OpenAI-compatible API keeps switching cost low, so a team can time it on its own prompts before committing. Its aicoolies score is 92/100 (Speed 98, Privacy 91, Developer Experience 90). This is a Documented score based on published sources, not an aicoolies measurement.