What Does Cerebras Inference Cost per Million Tokens?

Cerebras pricing at a glance

Cerebras is an inference provider that runs open-weight models on wafer-scale chips instead of GPU clusters, and it sells access two ways: a self-serve developer tier with published per-token rates, and an enterprise tier priced by quote. The published rates were checked on the official pricing page on October 26, 2026.

• gpt-oss-120b: $0.35 per million input tokens and $0.75 per million output tokens, listed at roughly 3,000 tokens/s.
• qwen-3.8-27b: $0.99 per million input tokens and $1.49 per million output tokens, listed at roughly 1,850 tokens/s.

Context also changes by tier. The model catalog gives gpt-oss-120b 65K tokens of context on the free trial and 131K on paid usage, while Qwen 3.8 27B gets 64K free and 128K paid. Enterprise buyers get custom weights, fine-tuning and reserved capacity by agreement, with the price set in a sales conversation.

What $5 of Cerebras credit actually buys

The signup offer is a one-time $5 promotional credit, and the mechanics matter more than the amount. Cerebras requires a valid payment method before granting it. The credit expires 30 days after activation, and the docs answer the obvious question directly: there is no permanently free tier.

At the gpt-oss-120b input rate, $5 covers about 14.3 million input tokens. On Qwen 3.8 27B it covers about 5.05 million. A job that sends one million input tokens and receives one million output tokens costs $1.10 on gpt-oss-120b, so the credit funds roughly four and a half runs of that size before it is gone.

When the credit is used up or expires, API and Playground access stop until you buy pay-as-you-go credits. Keys, projects and settings stay in place, and a purchase moves the account to the Developer tier with higher limits. That is a sensible shape for a two-week evaluation, but it is not a free tier you can leave a side project running on.

Cerebras vs Groq: same two models, different math

Groq publishes both of the same models on its own model card, which makes a like-for-like comparison possible without benchmarking anything myself. Groq lists gpt-oss-120b at 500 tokens/s with $0.15 input and $0.60 output per million tokens, and Qwen3.8-27B at 450 tokens/s with $0.80 input and $4.00 output.

On GPT OSS 120B, Cerebras is about six times faster and charges 2.3x more for input and 1.25x more for output. On Qwen 3.8 27B, Cerebras is about four times faster, charges 24% more for input, and charges 63% less for output. The output gap on Qwen is the biggest single difference between the two cards: one million input plus one million output tokens costs $4.80 on Groq and $2.48 on Cerebras.

That splits the decision cleanly. Prompt-heavy, short-answer work such as classification or routing is cheaper on Groq's GPT OSS 120B. Long-generation work, including agent transcripts and code, is cheaper on Cerebras for the Qwen model, and much faster on both. Rate limits also differ: Groq's developer plan lists 250K TPM and 1K RPM for these models, while Cerebras lists 1M uncached TPM and 1K RPM on gpt-oss-120b.

Two workloads, run through both price cards

Published rates only become a decision once they hit a realistic token mix. These are arithmetic on the two price cards above, not measured traffic.

A support assistant that processes 3 million input tokens and 600,000 output tokens a day costs $1.50 daily on Cerebras with gpt-oss-120b, or about $45 over 30 days. The same mix on Groq comes to $0.81 daily, or about $24.30. For a workload dominated by pasted context and short replies, Groq wins on cost and the speed advantage buys little.

A coding agent on Qwen 3.8 27B that takes 1 million input tokens and produces 2 million output tokens a day costs $3.97 daily on Cerebras, about $119 a month. The same volumes on Groq come to $8.80 daily, roughly $264 a month. Here the output-heavy shape flips the result: Cerebras is less than half the price and about four times faster, so the cheap option is also the fast one.

The rule that falls out: compare the output-to-input ratio of your own traffic before choosing a vendor. The two providers price output very differently, and that single ratio moves the bill more than the headline input rate does.

The trillion-parameter demo is not on the self-serve card

Cerebras published a technical post on October 22, 2026 explaining how it splits trillion-parameter models across multiple WSE-3T wafers, using Kimi K3 at roughly 3 trillion parameters as the worked example, with a stated target of 1,000 to 2,000+ tokens per second. A single WSE-3T wafer carries 44 GB of on-chip memory, and the post is explicit that trillion-parameter weights need terabytes, which is why the model spans several wafers.

That is a real capability, and it is also not what the pricing page sells. The developer tier lists two models, and the rate-limit docs say additional model families shown in the model chooser are available only through Dedicated Inference. So a team that wants Kimi K3 speed is buying reserved capacity through sales, not swiping a card for a $5 credit.

For a buyer, the practical read is simple: Cerebras' headline numbers describe its dedicated hardware deployments, while its published price list describes a narrow two-model catalog. Both are true. Confusing them is how teams end up budgeting for a frontier model that turns out to be quote-only.

Rate limits decide your real bill

Cerebras enforces two token buckets per organization. The uncached TPM limit covers cache misses and is the primary throughput constraint. The total TPM limit covers cached plus uncached tokens and defaults to three times the uncached figure. Both are enforced independently, and a 429 response names the bucket that was exceeded.

That design rewards prompt caching more than the price list suggests. With a 1M uncached TPM limit, the docs show a 3M total TPM ceiling, and at a 70% cache hit rate the account can process up to 3M tokens per minute while only 1M counts as uncached. For RAG services or chat apps resending the same system prompt, effective throughput can be several times the headline number.

One operational trap: Cerebras estimates a request's cost before running it, using your max_completion_tokens value as the output upper bound. Set that parameter far above what a response needs and you will hit 429s while your actual usage sits well under the limit. The docs recommend sizing it per workload. Developer-tier accounts also skip the hourly and daily token caps that apply to the free trial.

Who should pay for Cerebras, and who should not

The case for Cerebras is latency that changes what a product can do: voice agents that answer inside a conversational pause, inline code completion, live translation, anything where 3,000 tokens/s removes a spinner from the interface. Paying 2.3x Groq's input rate for gpt-oss-120b is defensible when the alternative is a worse user experience, and the $5 credit is enough to measure that on your own traffic before committing.

The case against is equally concrete. Batch summarisation, overnight document processing, offline evals and long-context retrieval do not care about tokens per second, and those workloads are cheaper elsewhere, including Groq for short GPT OSS 120B answers. Teams that need a frontier open-weight model in the multi-trillion class cannot self-serve it here at all.

A reasonable first test: put one real prompt against gpt-oss-120b, set max_completion_tokens to a realistic ceiling, and log time to last token plus token counts for a week. If the speed does not change your product, the price premium has no return. If it does, the two-model catalog is the limit to plan around.

At a glance

Published rates for the two models both providers sell, checked October 26, 2026

FAQ

Is Cerebras inference free?

No. Cerebras gives new accounts a one-time $5 credit after a verified payment method is added, and it expires 30 days after activation. When it runs out, API access stops until you buy pay-as-you-go credits.

How much does Cerebras charge per million tokens?

gpt-oss-120b costs $0.35 per million input tokens and $0.75 per million output tokens. qwen-3.8-27b costs $0.99 input and $1.49 output, per the official pricing page checked on October 26, 2026.

Can I run Kimi K3 or other trillion-parameter models on Cerebras?

Not on the self-serve tier. The developer catalog lists two models, and Cerebras' rate-limit docs state that additional model families are available only through Dedicated Inference, which is sold as reserved capacity.

Related reading

Qwen 3.8 27B on Cerebras vs Groq, Groq pricing, cheapest LLM API, cheapest GPU cloud for LLM inference

Sources

• Cerebras inference pricing (official)
• Cerebras rate limits and tier docs
• Cerebras model catalog
• Cerebras: deploying trillion-parameter models (Oct 22, 2026)
• Groq model card and published prices