Cerebras has no permanent free tier in 2026 — what looks like one is a one-time $5 promotional credit that expires 30 days after activation, after which your API and Playground access pause unless you buy pay-as-you-go credits.
If you found Cerebras through a “free LLM API” list, you are not alone — the offer reads like a free tier, and it is not one. This guide lays out what the trial is, what the $5 buys, which limit runs out first, and when to move to a provider you can keep using without paying.
Cerebras Has No Free Tier — It Has a $5 Trial
Direct answer: Cerebras does not offer a permanently free tier, only a time- and credit-bounded Free Trial of $5 in credits that expire 30 days after they’re granted, with no automatically renewing no-cost tier or per-model always-free allowance, per the official rate-limits FAQ.
A one-time credit you have to spend inside a window behaves very differently from a free tier you can leave running. Once the credit is used up or expires, the docs say you must buy credits to keep using Shared Inference, per the rate-limits FAQ.
The offer is newer than many lists assume. The change log dates the introduction of free trial credits for new accounts to 2026-07-16, alongside the rollout of uncached rate limits.
What You Actually Get When You Sign Up
Direct answer: At sign-up you get $5 in free credits after adding a verified payment method. The credits work across all public models and expire 30 days after grant. If you skip the payment method, Playground and API access stay inactive, per the rate-limits FAQ.
The payment method is what activates API access, so plan for it before you start. Credits apply across all public models, which makes the trial usable for a side-by-side comparison rather than a single-model demo, per the rate-limits FAQ.
There is no automatic billing trap. You are not automatically enrolled in paid usage, and access pauses when the credit expires or runs out unless you buy pay-as-you-go credits, per the pricing page. The pricing page counts the 30 days from activation, while the rate-limits docs count them from grant; the two anchors differ slightly, so treat the earlier date as your deadline. You can start at cloud.cerebras.ai.
Free Trial vs Developer Limits: The Real Numbers
Direct answer: Free Trial limits are far tighter than Developer limits. The trial caps both models at 5 RPM, 30K uncached TPM, 90K total TPM, 1M TPH, and 1M TPD. Developer tier raises throughput to 1M uncached TPM on gpt-oss-120b and removes the hourly and daily caps, per the rate-limits table.
The rows to watch are the last two. The trial has both an hourly and a daily token ceiling, and the Developer tier has neither, per the rate-limits table. That is a structural difference, not just a bigger number.
| Limit (per model) | Free Trial — gpt-oss-120b | Developer — gpt-oss-120b | Free Trial — qwen-3.8-27b | Developer — qwen-3.8-27b |
|---|---|---|---|---|
| RPM | 5 | 1,000 | 5 | 300 |
| Uncached TPM | 30,000 | 1,000,000 | 30,000 | 150,000 |
| Total TPM | 90,000 | 3,000,000 | 90,000 | 750,000 (temporarily raised from 450K) |
| Hourly cap | 1M TPH | none | 1M TPH | none |
| Daily cap | 1M TPD | none | 1M TPD | none |
| Extra limits | — | — | 2 images/request, 10 MiB payload | 10 images/request, 10 MiB payload |
Source: rate-limits docs; the qwen-3.8-27b Developer total TPM raise (450K to 750K, 2026-09-22) is in the change log.
Which Limit Runs Out First
Direct answer: In a single burst, the 30K uncached TPM bucket and the 5 RPM cap bind first. Over a full day, the credit runs out first at heavy use: at the 1M TPD ceiling, $5 lasts roughly 11 days on gpt-oss-120b and about 4.5 days on qwen-3.8-27b, per the rate-limits docs and pricing page.
Five requests per minute is the Free Trial RPM. The docs say a 429 names the bucket that was exceeded, and capacity replenishes continuously rather than resetting at fixed intervals. Space your calls out, and back off on 429s.
The bucket math matters more than the request count. Quota is estimated from your input plus max_completion_tokens before the request runs, so a 20K-token prompt with a 10K output cap can fill the 30K uncached TPM bucket in a single call, per the rate-limits docs.
The Dual-Bucket System: Why You Got a 429
Direct answer: Every model has two TPM buckets. The uncached bucket counts cache misses, and the total bucket counts uncached plus cached tokens and defaults to 3x uncached. Cached tokens don’t count against the uncached bucket, and 429 errors name the bucket you exceeded, per rate-limits docs.
That makes a 429 diagnosable. An uncached-TPM failure and a total-TPM failure call for different fixes, and the error message tells you which one you hit, per the rate-limits docs.
What $5 Actually Buys (Estimate)
Direct answer: At list prices, assuming 1,500 input plus 500 output tokens per request with no caching, $5 buys roughly 5,555 gpt-oss-120b requests at about $0.000900 each, or about 2,242 qwen-3.8-27b requests at about $0.002230 each, per Developer-tier pricing on the pricing page.
Label: ESTIMATE — not a measured figure. Assumptions: 1,500 input tokens and 500 output tokens per request, no caching, list prices, the full credit spent.
- gpt-oss-120b: 1,500 × $0.35/1M + 500 × $0.75/1M = $0.000900 per request, so $5 ≈ 5,555 requests (pricing)
- qwen-3.8-27b: 1,500 × $0.99/1M + 500 × $1.49/1M = $0.002230 per request, so $5 ≈ 2,242 requests (pricing)
- Daily-cap check: 1M TPD ÷ 2,000 tokens per request ≈ 500 requests/day. At that pace the credit runs out in about 11 days on gpt-oss-120b (5,555 ÷ 500) and about 4.5 days on qwen-3.8-27b (2,242 ÷ 500), both inside the 30-day window (rate-limits)
- RPM check: 5 requests per minute on the Free Trial (rate-limits)
- Bucket check: a 20K-token prompt with a 10K output cap counts 30K tokens against the 30K uncached TPM bucket before the request runs (rate-limits)
The Models on the Trial — and What Changed in 2026
Direct answer: The trial’s models are GPT OSS 120B (vendor-stated ~3,000 tokens/s, $0.35 input and $0.75 output per 1M tokens) and Qwen 3.8 27B (~1,850 tokens/s, $0.99 and $1.49), per the pricing page. Gemma 4 31B left public endpoints on September 3, 2026, per the change log.
You can read the full model pages for GPT OSS 120B and Qwen 3.8 27B. Both tokens/s figures are the vendor’s stated numbers, not measurements we ran ourselves, per the pricing page.
Two changes landed in September. On September 3, 2026, qwen-3.8-27b became available on public endpoints and gemma-4-31b was removed from them, per the change log. About three weeks later, on 2026-09-22, qwen-3.8-27b’s Developer total TPM was temporarily raised from 450K to 750K, also per the change log.
When to Switch to a Free-Forever Provider
Direct answer: Switch when your $5 credit is spent or the 30 days run out, since API and Playground access pause unless you buy pay-as-you-go credits, per the pricing page. Before that, compare our Groq free tier guide, Gemini API free tier guide, and OpenRouter free tier guide.
Treat day one of the trial as the start of a 30-day evaluation with a fixed end date, and have your fallback provider ready before the window closes.
If you want the landscape in one place, start with our best free AI API roundup for 2026, then branch into the Groq, Gemini, and OpenRouter guides for the providers you are likely to adopt.
The Bottom Line
Direct answer: Cerebras’s trial is a fast 30-day evaluation window, not a free tier. The pricing page says it is “for development, evaluation, and experimentation; not intended for production use,” per the pricing page, so plan your exit to a free-forever provider before day 30.
For a short test of whether very fast inference changes your product, $5 and 30 days is a low-cost way to find out. It is not a hosting plan: the pricing page’s footnote says the Developer tier is “not intended for production use”.
Track two numbers, not one: token consumption against the 1M TPD cap, and your remaining credit balance. Keep max_completion_tokens tight, since the estimate counts it before the request runs, per the rate-limits docs. Decide your destination provider before the credit expires.
Frequently Asked Questions
Does Cerebras have a permanently free tier?
No. The Free Trial is $5 in credits expiring 30 days after grant, with no automatically renewing no-cost tier or per-model always-free allowance. Once credits are used up or expire, you must purchase pay-as-you-go credits, per the rate-limits FAQ.
Do I need a payment method for the $5 credit?
Yes. New accounts receive $5 in credits after adding a verified payment method. If you skip it at sign-up, Playground and API access stay inactive until you add one, per the rate-limits FAQ.
What happens when the trial ends?
API and Playground access pause when the credit expires or is exhausted unless you buy pay-as-you-go credits. You are not automatically enrolled in paid usage, and your first PayGo purchase moves you to the Developer tier with higher limits, per the change log.
How This Guide Was Built
Direct answer: This guide was built from Cerebras’s official pricing page, rate-limits documentation, and change log, all fetched on 2026-10-21. We did not run the API hands-on, ran no benchmarks of our own, and all tokens/s figures are the vendor’s stated numbers.
Limits and bucket behavior come from the rate-limits docs. Prices, tier footnotes, and the $5 credit terms come from the pricing page. Dates, including the 2026-07-16 trial introduction, the September 3 model change, and the September 22 throughput raise, come from the change log. Arithmetic is labelled as an estimate with its assumptions stated inline.