Cerebras pricing

Cerebras' own pricing page for its inference API and coding subscriptions. Cerebras is not a model maker — it is a speed-differentiated inference provider that serves open models (Llama, Qwen and the like) on wafer-scale hardware cerebras-inference, so its pricing sells access + throughput, not a proprietary model. The page leads on the speed pitch ("20× faster than OpenAI and Anthropic") rather than on a per-token rate card — and notably does not publish per-million-token prices for individual models (corrected 2026-09-22: the page now carries a per-model rate card — see below), which is where most of the per-token market competes.

Per-model rate card (re-read 2026-09-22 — volatile)

The July page had no per-token prices. The September page lists two models with a speed figure and input/output prices per million tokens:

The speed numbers are Cerebras measuring its own service. The tier table now shows only two rows, Developer ("exploration workloads", self-serve pay-as-you-go with a free $5 credit) and Enterprise (production capacity, priority, custom weights, fine-tuning). The "from $10" Developer minimum and the "10× higher rate limits" line below were not on the page on 2026-09-22.

API tiers (July 2026 snapshot, as first read)

• Free trial — $5 in credits on signup; all models; community (Discord) support.
• Developer — self-serve, from $10; "10× higher rate limits than free"; priority processing.
• Enterprise — contact sales; highest rate limits, dedicated queue priority, custom model weights, fine-tuning/training, dedicated support.

Cerebras Code subscriptions (coding-agent product)

Flat monthly plans metered by a daily token bucket, not per-token billing:

• Pro — $50/mo — up to 24M tokens/day (the page frames this as ~$48/day of value); aimed at indie/weekend developers. Listed sold out.
• Max — $200/mo — up to 120M tokens/day (~$240/day of value); aimed at full-time and multi-agent use. Listed sold out.

Re-checked 2026-09-22 on cerebras.ai/code: both plans keep their prices and daily caps, now name GLM 4.7 as the model served, and are still listed sold out two months later. Both coding tiers being sold out is itself a signal: wafer-scale capacity is scarce, so Cerebras rations its cheapest fast-inference access — consistent with speed, not price, being the thing it actually sells.

Access routes

Also reachable via AWS, OpenRouter, Hugging Face, and Vercel (aggregators/resellers), so Cerebras throughput shows up inside other platforms' catalogs too.

Provenance, tier & cross-spoke

First-party vendor pricing page — T3 (the vendor stating its own prices, which is the first-party fact the page exists for), and a dated snapshot: tiers, dollar figures, and "sold out" status all drift; the pricing model (subscription token-buckets + per-use) is the durable part. Cross-spoke: the Cerebras Code subscriptions are a coding-agent product (an agentic-tooling-wiki concern) — noted here as context, not routed away; the wafer-scale mechanism and the company sit in llm-inference-wiki (cerebras-inference, cerebras-systems). See llm-api-pricing for how this speed-priced, subscription-bucket model contrasts with the per-token market.