Short answer. On GPT-OSS 120B, Groq lists $0.15 input and $0.60 output per million tokens at about 500 tokens per second, with a recurring free tier. Cerebras lists $0.35 and $0.75 at about 3,000 tokens per second, but its $5 credit expires after 30 days and requires a card.
Groq and Cerebras are often compared as "fast inference" vendors. They share an open model, an OpenAI-compatible request shape and a claim of unusually high throughput. They differ on price, free access and the commercial shape of the entry plan.
This is a 2027 planning guide built from vendor documentation checked on 20 November 2026. It is not a latency benchmark.
How we evaluated this
We weighted five criteria for a team choosing an inference vendor for a production feature.
- Cost per workload, 30%. Published input and output rates, applied to a stated token mix.
- Latency fit, 25%. Published throughput and the evidence available to compare it fairly.
- Free access, 15%. Whether the free option recurs, what it caps, and whether a card is required.
- Model coverage, 15%. Whether the model you need is production-grade or preview, and what else is in the catalogue.
- Operational fit, 15%. Context and output limits, rate limits and SDK compatibility.
We did not measure p50 or p95 latency from our own region. We did not run a quality evaluation. Vendor throughput figures are self-published and describe generation speed, not end-to-end response time. We did not verify enterprise contracts or committed capacity.
Side by side on the model both vendors publish
| Property | Groq | Cerebras |
|---|---|---|
| Model ID | openai/gpt-oss-120b | gpt-oss-120b |
| Input / 1M tokens | $0.15 | $0.35 |
| Output / 1M tokens | $0.60 | $0.75 |
| Published speed | ~500 tokens/sec | ~3,000 tokens/sec |
| Context window | 131,072 | 65,536 free; 131,072 paid |
| Max output | 65,536 | 32,768 free; 40,960 paid |
| Free access | Recurring free plan | $5 one-time credit, 30 days |
| Card required for free | No | Yes |
| Developer rate limit | 1,000 RPM; 250K TPM | 1,000 RPM; 500K uncached TPM |
Free tiers are not equivalent
Groq publishes a recurring free plan for GPT-OSS 120B: 30 requests per minute, 1,000 requests per day, 8K tokens per minute and 200K tokens per day. The limits apply at organisation level, so every key shares them. That is enough for development and a small internal tool, not for a launch.
Cerebras offers a one-time $5 promotional credit that expires 30 days after activation and requires a valid payment method. Its own pricing page says the Developer tier is for development, evaluation and experimentation, not production. The free trial limits on GPT-OSS 120B are 5 requests per minute, 30K uncached tokens per minute, 90K total tokens per minute and 1M tokens per day.
If you need a free tier that still works next month, Groq is the clearer choice. If you need a short, fast evaluation and have a card, Cerebras is viable.
Rate limits decide whether a cheap token price is usable. Groq's free tier caps GPT-OSS 120B at 8K tokens per minute, so a single 30K-token prompt will not fit even though the context window is 131K. Cerebras's free trial allows 30K uncached tokens per minute and 1M per day, with 5 requests per minute. Paid tiers raise these ceilings, but concurrency still bounds how many long requests can run at once. Size the plan against peak prompt length, not average monthly tokens.
Cost at volume
Assume a 70/30 input-to-output mix, list prices and no caching. At 10 million total tokens a month, Groq costs $2.85 and Cerebras $4.70. At 100 million, the same mix gives $28.50 and $47.00.
Groq, 10M tokens:
7M × $0.15/M + 3M × $0.60/M = $1.05 + $1.80 = $2.85
Cerebras, 10M tokens:
7M × $0.35/M + 3M × $0.75/M = $2.45 + $2.25 = $4.70
At 100M tokens: Groq $28.50; Cerebras $47.00
Groq is about 39% cheaper at this mix. That is a list-price result, not a claim about value. If Cerebras finishes a latency-bound task in a fraction of the wall-clock time, cost per completed task can favour it even at a higher token price. Model that separately.
Speed: what the numbers do and do not mean
Provider tokens-per-second figures describe generation throughput. Your p95 includes queueing, time to first token, network round-trip and output length. A 3,000 tokens-per-second engine with a two-second queue can feel slower than a 500 tokens-per-second engine with no queue.
Test the same prompt set from your own region, with your own concurrency, and record p50 and p95. Do not use a vendor speed claim as a capacity plan.
Model catalogue and production status
Groq's production table lists GPT-OSS 120B and GPT-OSS 20B, plus Whisper speech models. Qwen 3.8 27B appears in its preview table, not production, and is priced separately at $0.80 input and $4.00 output. Preview models can be withdrawn at short notice.
Cerebras publishes its supported models in its inference documentation. Check the specific model and its context limit before committing, because free and paid tiers can differ. Neither vendor is a full multi-model API: if you need vision, embeddings or a broad family of text models, expect a second provider.
Migration is a base URL change, not a rewrite
Both expose OpenAI-compatible chat completions. Keep the SDK, change the base URL and model ID, then re-test tools, structured outputs and streaming. The request body is the easy part; the differences show up in rate limits and error handling.
from openai import OpenAI
groq = OpenAI(api_key="GROQ_KEY",
base_url="https://api.groq.com/openai/v1")
cerebras = OpenAI(api_key="CEREBRAS_KEY",
base_url="https://api.cerebras.ai/v1")
When Groq is the better choice
Groq wins when token cost matters, when the team needs a free tier that persists, and when 500 tokens per second is already fast enough for the interface. It also wins when the workload depends on Whisper transcription or a broader production catalogue in the same account.
When Cerebras is the better choice
Cerebras wins when generation speed is the binding constraint: an interactive coding tool, a long reasoning trace, or a latency-sensitive agent loop where seconds of generation dominate the user experience. It also wins when a short evaluation fits the $5 credit. For production, plan the paid tier and its rate limits rather than the trial.
When neither is the right shape
Both are narrow catalogues with usage-based billing. If you need a flat monthly bill across several model families, embeddings, vision and private deployment, a multi-model API is a better fit. Plugsky Builder is $60 a month with no monthly token cap and no per-token overage; Scale is $120. Fair-use concurrency and RPM limits apply. See the pricing page.
What we did not verify
We did not verify real-world throughput, time to first token, uptime or error rates. We did not test the same prompts on both services, and no statement here is a benchmark. We did not verify enterprise pricing, committed capacity, regional availability beyond what the public pages state, or whether every preview model will remain available. We did not test tax treatment or currency conversion.
For Plugsky, we did not verify a SOC 2 certification. A SOC 2 Type II report is available under NDA; confirm the current scope and audit period.
Frequently asked questions
Is Groq or Cerebras cheaper?
For GPT-OSS 120B, Groq lists $0.15 input and $0.60 output per million tokens; Cerebras lists $0.35 and $0.75. At a 70/30 mix, Groq is about 39% cheaper on list price.
Which has a better free tier?
Groq has a recurring free plan with 30 requests per minute, 1,000 requests per day, 8K tokens per minute and 200K tokens per day for GPT-OSS 120B. Cerebras offers a one-time $5 credit that expires after 30 days and requires a card.
How fast are Groq and Cerebras?
Groq publishes about 500 tokens per second for GPT-OSS 120B; Cerebras publishes about 3,000. These are vendor generation figures, so measure p50 and p95 on your own prompts and region.
Can I use the OpenAI SDK with both?
Yes. Both publish OpenAI-compatible chat-completions endpoints. Keep the SDK, change the base URL and model ID, then re-test streaming, tools and structured outputs.
Which should I choose for production?
Choose Groq for lower token cost and a recurring free tier; choose Cerebras when generation speed is the constraint. For a flat bill across more models, compare Plugsky.
Do both support the same context length?
On GPT-OSS 120B, Groq lists 131,072 tokens of context. Cerebras lists 65,536 on its free tier and 131,072 on paid, with different maximum output lengths.
Sources
- Groq supported models and prices — checked 20 November 2026.
- Groq rate limits — checked 20 November 2026.
- Cerebras inference pricing — checked 20 November 2026.
- Cerebras GPT-OSS 120B model documentation — checked 20 November 2026.
Last checked 20 November 2026. Vendor prices and limits change; verify before committing. Not paid.