Cerebras

Key Terms

• Wafer-scale inference - Cerebras runs inference on CS-3 wafer-scale chips instead of GPUs, claiming 20x faster throughput than GPU-based providers. Source: Cerebras – Pricing
• Cerebras Code - a subscription coding product (separate from the API) providing token-based access to open-source models for IDE integrations and agentic workflows. Source: Cerebras – Introducing Cerebras Code
• Pay-per-token - the Developer API tier charges per token processed, similar to other inference providers. Source: Cerebras – Cerebras Inference Now Available Via Pay Per Token
• Inference provider - Cerebras does not train its own models. It runs third-party open-source models (GLM, GPT OSS, Llama, Qwen) on its hardware at high speed. Source: Cerebras – Pricing
• Preview model - a model available on the platform for evaluation only, not recommended for production workloads. GLM 4.7 and GPT OSS 120B are currently in preview. Source: Cerebras – Pricing

Latest Changes

First report for this supplier. All models, plans, and pricing are listed as current state.

• Deprecation (upcoming): Llama 3.1 8B and Qwen 3 235B Instruct will be deprecated May 27, 2026.
• Feature added: Blog posts on multi-agent workflows and Figma multi-agent integration (April 16).
• Feature added: MCP vs. CLI speed debate blog post (April 6).
• Plan change: Cerebras Code Pro ($50/mo) and Max ($200/mo) remain sold out with no restock ETA.
• Feature added: Cerebras available through AWS Marketplace (March 13).

Plans

Inference API Tiers

Cerebras Code (Coding Agent Product)

Source: Cerebras – Pricing

API Pricing

Developer API Per-Token Pricing

Implied Code product pricing: Cerebras Code Pro provides 24M tokens/day at $50/month, which works out to roughly $0.07/1M tokens if fully utilized (720M tokens/month). This is significantly below API rates, suggesting the Code product bundles a subsidized allocation.

Source: Cerebras – Pricing

Model Performance / Benchmarks

Cerebras does not train its own models and does not publish benchmark scores. It hosts third-party open-weight models on wafer-scale chips. Key performance claim:

• 1,000 to 3,000 tokens/second generation speed (confirmed by published pricing page)
• 20x faster throughput than GPU-based providers (claimed)

The practical impact for agentic coding, where thinking time often dominates generation speed, is debated in the community. Speed advantage is most relevant for high-throughput tasks where token generation is the bottleneck.

Source: Cerebras – Pricing

Latest News

Lessons Learned from Building Multi-Agent Workflows (April 16, 2026)

Blog post sharing engineering insights from deploying multi-agent systems on Cerebras infrastructure. Covers practical challenges in orchestrating multiple AI agents for production workloads.

Source: Cerebras – Lessons Learned From Building Multi Agent Workflows

Figma MultiAgents (April 16, 2026)

Blog post about a multi-agent integration with Figma, demonstrating how fast inference enables real-time collaborative AI-assisted design workflows.

Source: Cerebras – Figma Multiagents

The Debate of MCP vs. CLI Centers on Speed (April 6, 2026)

Blog post arguing that the choice between MCP (Model Context Protocol) servers and CLI-based agent interfaces comes down to inference speed. Cerebras positions its speed advantage as critical for CLI-based agentic workflows where round-trip latency compounds.

Source: Cerebras – Mcpvscli

Cerebras Coming to AWS (March 13, 2026)

Cerebras announced availability through AWS Marketplace, allowing customers to access Cerebras inference through AWS with flexible pricing and low-latency workloads. This is a distribution expansion, not a new product.

Source: Cerebras – Cerebras Is Coming To Aws

Introducing Cerebras Code (August 1, 2025)

Cerebras Code launched as a subscription coding product with Pro ($50/month) and Max ($200/month) tiers. Both tiers are currently sold out.

Source: Cerebras – Introducing Cerebras Code

Qwen3 Coder 480B Live on Cerebras (August 1, 2025)

Qwen3 Coder 480B model made available on Cerebras inference, a key coding-focused model in their lineup.

Source: Cerebras – Qwen3 Coder 480B Is Live On Cerebras

Community Signals

• Cerebras Code Pro and Max tiers have been sold out for an extended period. No ETA for reopening has been communicated. This limits access to the most cost-effective coding plans.
• Cerebras is frequently mentioned in discussions about inference speed, with community benchmarks confirming 10-20x faster token generation compared to GPU-based providers. However, the practical impact for coding tasks (where thinking time dominates over generation speed) is debated.
• The OpenAI Codex Spark partnership (February 2026) generated significant interest, as it positions Cerebras as the fast-inference backend for an OpenAI product.

Enterprise Readiness

Transparency Gaps