Cerebras Pricing: API Costs and the Production Catch

Cerebras pricing starts at $0.35 per million input tokens and $0.75 per million output tokens for GPT OSS 120B, but the Developer tier is explicitly not intended for production use. A worked agent workload below costs $53.75 a month at those rates; launching it for customers calls for a separate production arrangement, not just more credits.

Cerebras Inference is a hosted API for running AI models. The buying decision turns on three things: cost per completed task, time to finish that task, and whether the access terms fit your deployment. Start with GPT OSS 120B for a text-agent evaluation; move to a production discussion before making customer traffic depend on it.

Prices verified against Cerebras's pricing page, October 2026, with the live check performed on 20 October. All amounts below are in US dollars.

Cerebras Pricing at a Glance

The current pricing page names Developer and Enterprise. The promotional credit is a way to evaluate the service, not a third permanent free subscription.

Enterprise also offers enhanced latency options; customized model weights and fine-tuning or training services are available by agreement. Developer support is through the community Discord. These are the distinctions in the live tier comparison.

The decision rule is workload intent. An inexpensive customer-facing agent still needs suitable production terms. A large evaluation bill does not, by itself, establish a need for a particular Enterprise contract.

Cerebras API Pricing: Each Listed Model and Speed

GPT OSS 120B is the cheaper of the two models in the current Developer rate card. Input tokens are the pieces of text you send; output tokens are what the model generates. The rates apply separately, so a single blended price is only meaningful after you choose an input-to-output mix.

Cerebras Models List

OpenAI GPT OSS 120B: $0.35 per million input tokens, $0.75 per million output tokens, and a Cerebras-listed speed of approximately 3,000 tokens per second. Its model documentation lists text input and output, reasoning, streaming, structured outputs and tool calling. That makes it the first candidate here for a text-based agent, subject to your own task-quality checks.

Qwen 3.8 27B: $0.99 per million input tokens, $1.49 per million output tokens, and a Cerebras-listed speed of approximately 1,850 tokens per second. Its model documentation includes image input and strict tool calling. Image input requires Chat Completions and base64-encoded PNG or JPEG data; an ordinary external image URL is not supported.

For smaller units, the GPT OSS rates are $0.00035 input and $0.00075 output per 1,000 tokens. Qwen's are $0.00099 and $0.00149. Those are unit conversions, not separate plans.

These rates cover inference, not a subscription to your coding editor. If your editor or agent service has its own fee, add that separately to the API budget.

The speeds above are vendor figures, not measurements from this article. A coding agent also waits for prompt processing, reasoning, tool execution, network calls and retries. Fast generation can shorten a long answer while barely changing a task dominated by running tests or querying a slow service.

What Does an Agent Workload Cost per Month?

A text agent handling 10,000 tasks per month costs $53.75 on GPT OSS 120B under the assumptions below. This is an illustrative evaluation budget at public Developer rates, not an Enterprise quote or measured production bill.

Assume each task takes five model calls, and each call averages 2,000 input tokens and 500 billable output tokens. The output allowance must include any charged reasoning, not just the final text shown to the user. Assume no cache savings, no retries and no separate tool charges.

The monthly token totals are:

• Input: 10,000 tasks × 5 calls × 2,000 tokens = 100 million tokens.
• Output: 10,000 tasks × 5 calls × 500 tokens = 25 million tokens.

Using the listed GPT OSS 120B rates, input costs 100 × $0.35 = $35, and output costs 25 × $0.75 = $18.75. Total: $53.75 per month, or $0.005375 per task before any promotional credit.

The same token volume on Qwen 3.8 27B costs $99 input + $37.25 output = $136.25 per month. That is an extra $82.50 for this scenario. It is worth paying only if the different model or image-input capability improves the task enough to justify it; identical token counts do not imply identical results.

For your own forecast, multiply monthly input millions by the input rate, then add monthly output millions multiplied by the output rate. For unit economics, divide the resulting bill by successfully completed tasks, including the spend consumed by failed attempts.

Is Cerebras Free, and Where Does the Trial Stop?

Cerebras offers a bounded trial, not an allowance that renews forever. Its trial documentation specifies a one-time $5 credit, activated by adding a valid payment method, which expires 30 days after activation. Adding the payment method does not automatically enroll you in paid usage or charge you.

When the credit is exhausted or expires, API and Playground access pauses unless you separately purchase PayGo credits. Skipping the payment method leaves that access inactive. The promotional credit is not cash and cannot be transferred.

The documented trial defaults for both listed models are 5 requests per minute, 30,000 uncached tokens per minute, 90,000 total tokens per minute, and 1 million tokens per hour and per day. Your account's Limits screen is the authority for your actual allowance. The token-per-minute ceilings count work, while the request ceiling counts calls; either can stop a test before the credit balance runs out.

Who never needs to pay? Someone who finishes a small evaluation within the credit, time and rate limits, then stops using the service. A developer wanting an ongoing personal coding assistant should budget for paid usage. The trial is useful for deciding whether to continue, not for forecasting a zero-cost service indefinitely.

The Limits That Matter Before You Ship

The first practical wall is burst capacity: a low monthly bill does not mean the API can accept every simultaneous request. Cerebras documents separate limits for uncached tokens, which require new processing, and total tokens, which also include cache hits.

The current Developer defaults are:

• GPT OSS 120B: 1,000 requests per minute, 1 million uncached tokens per minute and 3 million total tokens per minute.
• Qwen 3.8 27B: 300 requests per minute, 150,000 uncached tokens per minute and 750,000 total tokens per minute. The documentation marks the total-token allowance as a temporary increase from 450,000.

Developer has no hourly or daily token cap, but these per-minute limits still apply. Organization and project limits can constrain the same workload, so dividing your monthly total by the number of days is not a capacity check.

There is another easy-to-miss limit: the service estimates output demand before running a request. An unnecessarily large max_completion_tokens setting can trigger rate limiting before the model produces anything. Set it to an appropriate ceiling for the task, then inspect the returned usage and quota errors. Purchasing more credits does not remove these controls.

For coding agents, reliability matters as much as capacity. The GPT OSS model notes warn that it can propose tools that were not supplied. Validate tool names and arguments before execution. A fast incorrect action is still an incorrect action.

Enterprise and the Four Partner Routes

Production use changes the purchasing conversation even when token consumption is small. There is no public Enterprise rate from which to calculate a universal Developer-to-Enterprise spending break-even. Ask for a quote using your model, expected token mix, peak concurrency and support requirements.

Cerebras's pricing page names four partner routes: AWS Marketplace, OpenRouter, Hugging Face and Vercel. Consider the purchasing route you already operate, but check its own rate card, model availability, production terms and support responsibility before moving traffic. A named partnership does not establish that its final bill matches the direct Developer rate.

The direct rate card is usage-based; it does not publish an annual-versus-monthly subscription discount or an Enterprise overage schedule. For a quote, get minimum commitments, excess-usage charges, capacity guarantees and exit terms in writing. For prepaid usage, check paid-credit expiry and refund terms before buying a balance. The promotional credit's expiry should not be assumed to describe purchased credits.

Cerebras vs Groq and a GPU Cloud API

Groq costs less for the same uncached GPT OSS 120B token mix, while Cerebras advertises faster generation; the Groq pricing guide covers that alternative's plan and capacity tradeoffs. Groq is another hosted inference provider. Its official model page lists $0.15 per million input tokens, $0.60 per million output tokens and approximately 500 tokens per second.

At the example's 100 million input and 25 million output tokens, Groq's token bill is $30 per month. Cerebras's $53.75 Developer estimate is $23.75 more, or 0.2375 cents more per task. These are public-rate calculations, not a comparison of negotiated production contracts. The vendor speed figures also do not establish how much faster your complete agent would run.

Fireworks AI, a hosted model provider with serverless and dedicated deployment options, gives a second comparison. Its Standard GPT OSS 120B rates are also $0.15 per million input and $0.60 per million output tokens, producing the same $30 monthly estimate without caching. Priority serving is priced separately; this comparison uses Standard throughout.

For either alternative, compare the same reasoning settings, output limits and task-quality threshold. Do not compare a low-cost setting that fails your task with a more expensive setting that completes it, or a rented GPU's hourly sticker price with an API bill that already includes serving infrastructure.

The useful economic question is whether the measured latency improvement is worth the premium per successful task. For a coding tool used interactively, waiting time may matter. For a background agent already dominated by external tool calls, cheaper tokens may be the better purchase.

What to Check Before Switching

Evaluate with a representative task set before migrating the application. A short answer benchmark alone will not reveal whether a coding agent follows tool schemas or completes a long edit correctly.

1. 1

   Measure the existing task

   Record input and output usage across every model call, retries, successful completions and total elapsed time. Separate model time from time spent in tools.

2. 2

   Replay the task on the candidate route

   Use equivalent reasoning and output settings. Check tool-call validity and task correctness before judging speed. Include slow and failed tasks in the comparison.

3. 3

   Price the deployment you will run

   Apply the chosen provider's rates to the observed usage, then confirm capacity and production terms. For direct Cerebras production access, obtain the Enterprise quote before switching customer traffic.

Next week, replay a representative slice of your agent workload and bring the resulting usage and concurrency figures to the purchasing discussion. For a broader shortlist, see the real-time inference platforms for agents.

Cerebras Pricing FAQ

Can I use Cerebras for free?

Yes, within the one-time promotional trial after adding a valid payment method. It is time-limited and credit-limited; there is no automatically renewing free allowance. Finishing an evaluation and stopping can cost nothing.

Is Cerebras better than nvidia?

There is no universal winner for an agent. NVIDIA is a hardware vendor, while this purchase is a hosted inference service. Compare suitable API deployments on task quality, total latency, capacity and the full bill; the chip architecture alone does not decide those outcomes.

Is it worth buying Cerebras?

For inference access, it is worth evaluating when model generation is a meaningful part of your application's waiting time. Pay for the route only after its task results, measured latency and operating terms justify the total cost.

Who are Cerebras main competitors?

For this fast-inference API decision, shortlist Groq and a GPU cloud inference provider such as Fireworks AI. Compare the same model and workload before treating different providers' starting prices as interchangeable.

How much does a Cerebras chip cost?

Cerebras hardware is sold by quote under its Terms of Sale; the company's stock is a separate investment question.

Does Cerebras publish a student discount or refund policy on its pricing page?

The current pricing page publishes neither a student-specific rate nor a general paid-credit refund promise. Confirm the applicable credit-purchase terms before paying; do not treat promotional credit as refundable cash.

Use the AI Business Workflow Audit Checklist to map task volume, failure handling and operating costs before choosing an API. Subscribe to receive the checklist.