Short answer: Cerebras shared inference costs $0.35 input and $0.75 output per million tokens for gpt-oss-120b, and $0.99/$1.49 for qwen-3.8-27b. The paid tier raises context from 65K to 131K on GPT-OSS and from 64K to 128K on Qwen, and raises rate limits to 1,000 RPM and 500K uncached tokens per minute on GPT-OSS. Cached input and batch pricing are not published, so the only lever is model choice and prompt size.
Cerebras prices per million tokens
| Model | Input | Output | Cached input | Context (free / paid) | Max output (free / paid) |
|---|---|---|---|---|---|
gpt-oss-120b | $0.35 | $0.75 | Not published | 65K / 131K | 32K / 40K |
qwen-3.8-27b | $0.99 | $1.49 | Not published | 64K / 128K | 32K / 40K |
USD per 1 million tokens, retrieved 19 November 2026 from Cerebras Inference model documentation and the Cerebras pricing page. Cerebras publishes no cached-input rate and no batch rate for shared inference. Dedicated inference and enterprise capacity are “available by agreement” and quoted through sales, so no public per-token price exists for those tiers.
Both prices are flat within the context window. A 100K-token request on the paid tier pays the same rate per token as a 1K-token request. The context window itself is the tier boundary: on the free trial, GPT-OSS stops at 65K tokens and Qwen at 64K, so a long-document job does not become expensive on the free tier, it becomes impossible.
The free trial is a credit, not a free tier
Cerebras gives new accounts a one-time $5 promotional credit after a valid payment method is added. The credit expires 30 days after activation. API and Playground access pause when it is used up or expires unless you purchase pay-as-you-go credits. Cerebras states plainly that it does not offer a no-cost tier that renews automatically or a per-model always-free allowance.
The trial's rate limits are restrictive for GPT-OSS: 5 requests per minute, 30K uncached tokens per minute, 90K total tokens per minute and 1M tokens per day. Qwen is far more usable on the trial at 300 RPM, 150K uncached tokens per minute and 750K total tokens per minute, with no hourly or daily token cap. If you are evaluating Cerebras for a latency-sensitive product, test with Qwen first; the GPT-OSS trial lane will throttle a realistic load test before the model does.
At $5, the credit buys about 14.3 million GPT-OSS input tokens or 6.7 million output tokens if spent on one side only. That is enough for a benchmark, not a pilot. Budget for a paid top-up before you promise a stakeholder a production trial.
What changes the Cerebras bill
Prompt caching helps throughput, not price
Cerebras supports prompt caching and enforces two token buckets: an uncached TPM limit and a total TPM limit set at 3× the uncached limit (temporarily 5× for Qwen 3.8 27B). Cached tokens do not count against the uncached bucket, so a high cache-hit rate lets the same purchased throughput serve far more total tokens. The documented example is a 1M uncached TPM limit with a 3M total limit: at a 70% hit rate you can process up to 3M total tokens per minute.
What Cerebras does not publish is a lower price for cached input. Unlike Groq's explicit 50% cache discount or Qwen's implicit-cache rate, the Cerebras model page lists one input rate. Treat caching as a capacity lever, not a cost lever, and do not put a cache saving in a budget without confirming it in a current quote.
No batch or long-context surcharge published
There is no published batch endpoint discount and no long-context multiplier. The bill is input tokens × input rate plus output tokens × output rate, with the context ceiling determined by tier. This makes Cerebras one of the simplest bills to forecast once you know your token counts, and one of the least flexible if you need a discount for offline work.
Output and model mix
Output costs 2.14× input on GPT-OSS and 1.5× on Qwen. That is a milder output penalty than Groq's 4× on GPT-OSS 120B, so Cerebras is relatively better for generation-heavy workloads than its headline input rate suggests. The absolute rates are still higher: a 10M-input, 2M-output job costs $5.00 on Cerebras GPT-OSS versus $2.70 on Groq at list price, before Groq's cache or batch discounts.
Qwen 3.8 27B is the more expensive model per token and the slower one at about 1,850 tokens per second, but it is multimodal and supports image inputs. Choose it for vision or Qwen-specific behaviour, not for cost.
Rate limits are the hidden upgrade trigger
The paid tier's 1,000 RPM and 500K uncached TPM on GPT-OSS are generous for most applications, but the free trial's 5 RPM is not. A team that outgrows the trial is not choosing between $0 and $0.35 per million; it is choosing between a throttled prototype and a paid tier. Size the jump from your p95 tokens per request, not from the monthly token total alone.
Worked example: 10M input, 2M output
| Model | Input cost | Output cost | Total |
|---|---|---|---|
gpt-oss-120b | 10M × $0.35 = $3.50 | 2M × $0.75 = $1.50 | $5.00 |
qwen-3.8-27b | 10M × $0.99 = $9.90 | 2M × $1.49 = $2.98 | $12.88 |
The same token mix on Groq GPT-OSS 120B costs $2.70 at list price, or $2.175 with a 70% cache hit. Cerebras is 1.85× the uncached Groq bill and 2.3× the cached one. In exchange, Cerebras advertises about 3,000 tokens per second versus about 500 on Groq, a roughly 6× speed advantage. Whether that trade is worth it depends entirely on whether latency changes revenue, completion rate or infrastructure cost in your product.
When a flat plan wins
A flat plan wins when your latency target can be met without Cerebras. At $5 per 10M input plus 2M output, $60/month buys about 120 million input and 24 million output tokens at that mix. Above that volume, a $60 unlimited-fair-usage plan is cheaper unless the Cerebras speed is load-bearing. Below it, the per-token bill is small enough that a subscription may be the more expensive way to buy the same tokens.
The stronger case for a flat plan is flexibility. Cerebras shared inference offers two models. A 39-model OpenAI-compatible endpoint lets you route easy requests to a cheap model and reserve the fast, expensive lane for the few requests where latency matters. That routing decision usually saves more than a cache tweak, because most traffic does not need 3,000 tokens per second.
Plugsky's plans start at $8/month for Hobby, with Builder at $60/month for unlimited fair usage, 256K input, 32K output, 10 concurrent requests and 600 RPM. Plugsky does not claim Cerebras-class tokens per second. Choose Cerebras when the speed is the product; choose a flat plan when it is not.
How to verify these numbers
- Open
cerebras.ai/pricingfor the tier structure, then the model page atinference-docs.cerebras.ai/models/overviewfor the per-token rate. - Confirm whether the context and max-output figures you need are on the free trial or paid tier.
- Check the rate-limit page for the exact uncached and total TPM values; they differ by model.
- Ask sales for a dedicated quote if you need reserved capacity or a volume discount; the public page does not publish one.
- Re-check after any model addition. Some model families are available only through Dedicated Inference, not shared inference.
Use the LLM API cost calculator to put Cerebras and a standard-speed provider on the same token mix, and the rate-limit calculator to see whether the 5 RPM trial or the 30K uncached TPM cap binds first. The free API tracker lists which providers have a renewable free tier, and the cheapest LLM API guide ranks by workload rather than speed.
FAQ
How much does Cerebras charge per million tokens?
On 19 November 2026 Cerebras listed gpt-oss-120b at $0.35 input and $0.75 output per million tokens, and qwen-3.8-27b at $0.99 input and $1.49 output. Both are shared-inference pay-as-you-go rates; dedicated capacity is quoted separately.
Does Cerebras have a free tier?
No permanent free tier. New accounts can receive a one-time $5 promotional credit after adding a valid payment method. The credit expires 30 days after activation, and API and Playground access pause when it is used up or expires.
Is Cerebras prompt caching cheaper?
No cached-input discount is published. Cerebras prompt caching reduces rate-limit pressure because cached tokens do not count toward the uncached TPM bucket, but the published input price is the same for cached and uncached tokens.
Why is Cerebras more expensive than Groq for GPT-OSS?
Cerebras lists gpt-oss-120b at $0.35 input and $0.75 output, against Groq at $0.15 and $0.60. The premium buys roughly 3,000 tokens per second on Cerebras versus about 500 on Groq. You are paying for latency, not for a different model.
Most requests do not need 3,000 tokens per second
Route the latency-critical few to a fast lane and the rest to a flat-rate API. Plugsky is OpenAI-compatible and starts free with no card.
Start Free → See plans