Groq vs Cerebras: Fast Inference Compared (2026)
November 18, 2026

Most Groq vs Cerebras comparisons still describe long lists of Llama models on both sides. That picture is out of date. As of November 2026, the self-serve catalogs at both companies have shrunk to nearly the same short list, and the two flagship chat models are identical: OpenAI's open-weight GPT OSS 120B and Alibaba's Qwen 3.8 27B.
So the choice is no longer "who hosts the model I want". It comes down to four things:
- Price per token. Groq lists lower prices on GPT OSS 120B. On Qwen 3.8 27B, Groq is cheaper on input and Cerebras is cheaper on output.
- Speed. Both sell speed, and both publish their own numbers. Independent measurements favour Cerebras on the shared flagship model, with caveats.
- What else you need. Groq also serves speech-to-text, text-to-speech and a smaller 20B model. Cerebras's shared tier is two chat models.
- How you start. Groq has a standing Free tier. Cerebras gives a one-time trial credit.
Everything below was checked on Groq's model docs, Cerebras's model catalog, and each company's pricing and press pages on 18 November 2026.
Quick comparison
| Groq | Cerebras | |
|---|---|---|
| Hardware | LPU (Language Processing Unit) | Wafer-Scale Engine (WSE) |
| Self-serve chat models | GPT OSS 120B, GPT OSS 20B, Qwen 3.8 27B (preview) | GPT OSS 120B, Qwen 3.8 27B |
| Speech models | Whisper Large v3 and v3 Turbo, Orpheus text-to-speech (preview) | None on the shared tier |
| Starting tier | Free tier with rate limits | One-time $5 credit, 30 days, payment method required |
| Paid self-serve | Developer tier, pay per token | Developer tier, pay as you go |
| Batch | 50% discount, Developer tier | Private preview |
| More models | Enterprise contract | Dedicated Inference |
| API | OpenAI-compatible | OpenAI-compatible |
| Company status | Says it operates as an independent company; licensed its technology to Nvidia in December 2025 | Public, Nasdaq: CBRS |
The model catalogs, as of November 2026
Groq
Groq's models page splits the catalog into production and preview.
Production models:
openai/gpt-oss-120bopenai/gpt-oss-20bwhisper-large-v3andwhisper-large-v3-turbofor speech-to-text
Preview models, which Groq says are for evaluation and may be discontinued at short notice:
qwen/qwen3.8-27bopenai/gpt-oss-safeguard-20bfor content moderation- Two Llama Prompt Guard 2 models
- Canopy Labs Orpheus text-to-speech, in English and Saudi Arabic
A few more rows are tagged Enterprise with "Contact Sales" in place of a price: Llama 3.1 8B, Llama 3.3 70B, Meta's Muse Spark Open and MiniMax M2.7. You can't use those on a self-serve account.
That enterprise tag is the trap. Groq's deprecations page shows it shut down llama-3.1-8b-instant and llama-3.3-70b-versatile for free and developer-tier usage on 16 August 2026, and pointed users to the GPT OSS models. It also retired qwen/qwen3.6-27b on 14 September and its groq/compound agent systems on 21 September. If a tutorial tells you to call Llama 3.3 70B on Groq with a personal key, it predates all of this.
Cerebras
Cerebras's Shared Inference catalog lists exactly two models:
gpt-oss-120bqwen-3.8-27b
Its deprecations page tells a similar story of consolidation. llama-3.3-70b was deprecated in February 2026, llama3.1-8b in May, and zai-glm-4.7 in August. Gemma 4 31B left Shared Inference on 3 September 2026 and is now available only on Dedicated Inference.
One detail is easy to miss: context length depends on your tier. Cerebras lists GPT OSS 120B at 65k tokens of context on the trial tier and 131k on paid. Groq lists the same model at 131,072 tokens.
What this means
On a credit card and an API key, you get the same two headline models from either company. Groq adds a small fast model and audio. Cerebras adds nothing on the shared tier but opens a wider range of model families, plus your own fine-tuned weights, through Dedicated Inference, which is reserved capacity sold by agreement.
If you need a broad menu of open models on demand, neither is the right first stop any more. See our roundups of Together AI alternatives and Fireworks AI alternatives for hosts with bigger catalogs.
Pricing and starting tiers
Both charge per million tokens. These are the list prices on each vendor's own page on 18 November 2026:
| Model | Groq input / output | Cerebras input / output |
|---|---|---|
| GPT OSS 120B | $0.15 / $0.60 | $0.35 / $0.75 |
| Qwen 3.8 27B | $0.80 / $4.00 | $0.99 / $1.49 |
| GPT OSS 20B | $0.075 / $0.30 | Not offered |
Groq's figures come from its models page. Cerebras's come from its pricing page.
On GPT OSS 120B, Groq is cheaper on both sides. Qwen 3.8 27B is more interesting: Cerebras charges a little more for input and far less for output. A workload that writes long answers will cost less there. One that reads long documents and returns short answers will cost less on Groq.
Getting started differs more than the token prices do.
- Groq has a standing Free tier with low rate limits. Its billing FAQ says upgrading to the Developer tier needs a payment method, has no upfront charge, and unlocks higher limits, Flex processing, spend limits and batch jobs. Batch processing is listed at a 50% discount.
- Cerebras has no standing no-cost tier. Its pricing page describes a one-time $5 promotional credit after you add a payment method, expiring 30 days after activation. After that, access pauses unless you buy pay-as-you-go credits. The same page labels the Developer tier as "not intended for production use" and points production workloads to Enterprise. Its Batch API is in private preview.
Rate limits follow the same pattern. Cerebras's rate limits page lists GPT OSS 120B at 5 requests per minute on trial credits and 1,000 per minute on pay as you go. Groq's limits vary by model and plan, and its docs tell you to read the exact numbers for your organization in the console.
Speed: what you can and can't trust
Speed is the reason both companies exist, so it is also where marketing is loudest.
Each vendor's own number. Groq's models page lists GPT OSS 120B at about 500 tokens per second. Cerebras's catalog lists the same model at about 3,000. Those figures are measured by the companies themselves, on their own setups, and Cerebras's site carries a footer noting that performance comparisons vary by workload, configuration, date and model.
An independent number. Artificial Analysis benchmarks hosted endpoints and publishes a page per provider. On 18 November 2026:
- Its Cerebras page reported GPT OSS 120B (high reasoning) at 1.0 minutes per Intelligence Index task.
- Its Groq page reported the same model and setting at 3.2 minutes per task.
Time per task is output tokens divided by end-to-end output speed, so lower is faster. On that measure Cerebras finished about three times sooner.
That is a real gap, and it is still one benchmark on one model. Your result depends on prompt length, output length, reasoning effort, time of day, which service tier you're on, and how far you are from the data center. Qwen 3.8 27B may not rank the same way.
How to test it yourself:
- Take 20 real prompts from your app, not toy questions.
- Send them to both APIs with the same model, the same reasoning setting and the same output cap.
- Record time to first token and total time, at the hours your users are active.
- Multiply by your real token mix to get a monthly cost on each side.
Both APIs are OpenAI-compatible, so this is an afternoon of work.
Hardware: LPU vs Wafer-Scale Engine
You never touch the chips, but they explain the trade-offs.
Groq's LPU. Groq describes the LPU as a processor it designed from scratch for inference. In its telling, the design is software-first, built around a "programmable assembly line" that moves data between chips without routers or controllers, and deterministic, so each step takes a known amount of time. Many LPUs work together to run one model.
Cerebras's Wafer-Scale Engine. Cerebras goes the other way and makes one enormous chip. Its chip page describes the current WSE-3 Turbo as a single wafer-sized processor with 4 trillion transistors and 900,000 AI cores across 46,225 square millimeters. The pitch, in Cerebras's words, is that putting compute and memory on one wafer reduces data movement and chip-to-chip communication. In August 2026 it announced the CS-4, a rack-scale system built from three of these wafers.
Both descriptions are vendor claims about their own silicon. What matters to you is the output of the test in the previous section.
Company status: what each one says about itself
Infrastructure choices last years, so it is fair to ask who you're buying from. Aggregator sites get this part wrong more than any other.
Groq
On 24 December 2025, Groq announced "a non-exclusive licensing agreement with Nvidia for Groq's inference technology". The same announcement said founder Jonathan Ross, president Sunny Madra and other team members would join Nvidia, and that "Groq will continue to operate as an independent company".
Groq does not call this an acquisition, and neither should you. GroqCloud kept running, and the company has made further announcements since:
- In June 2026 it announced $650 million in new capital and a strategy focused on being an inference cloud.
- In August 2026 it announced a $350 million Series A, described as subject to customary closing conditions, with planned participation from Nvidia.
- It joined the NVIDIA Cloud Partner program, and says it will be among the first to deploy NVIDIA's Groq 3 LPX systems.
In practical terms: the founder and president moved to Nvidia, the API stayed, and the roadmap now includes Nvidia hardware.
Cerebras
Cerebras is a public company. Its own press releases identify it as Cerebras Systems (NASDAQ: CBRS), and its newsroom lists third-quarter 2026 financial results.
On the product side, the pricing page lists partner access through AWS Marketplace, OpenRouter, Hugging Face and Vercel. In October 2026, Cerebras and Oracle announced plans to offer Cerebras inference on Oracle Cloud Infrastructure, planned for calendar year 2027.
Neither status makes one vendor safer by default. A public company publishes its finances. Groq has announced new funding, part of it still subject to closing, and a deepening Nvidia partnership. Both have retired popular models from their self-serve tiers this year, which is the risk that will affect your code.
Which one should you pick?
Pick Groq if:
- You want to prototype on a standing Free tier before you add a payment method.
- Input-heavy workloads dominate, or you mainly use GPT OSS 120B and price matters most.
- You need speech-to-text or text-to-speech from the same provider.
- You want a 50% batch discount today.
Pick Cerebras if:
- Your own test confirms the speed gap and latency is what your product sells.
- You run Qwen 3.8 27B with long outputs.
- You plan to move to reserved capacity or deploy custom weights.
Use both if you can. With two OpenAI-compatible endpoints serving the same models, a fallback from one to the other is a small amount of code, and it protects you from the next deprecation notice.
For other fast-inference hosts, we keep longer lists of Groq alternatives and Cerebras alternatives.
If you'd rather not wire up an API at all
Comparing inference providers assumes you're building the app yourself. Plenty of people just want the finished tool.
If you'd rather run a tool someone already built than set one up yourself, that's where Taku comes in. Taku Apps and Sites, built by the best AI minds. Find one, run it, and keep it in your Mac Dock. Taku is in Beta, and the Mac app is available now.
FAQ
Did Nvidia acquire Groq?
Groq's own announcement describes a non-exclusive licensing agreement for its inference technology, with its founder and other staff joining Nvidia, and says Groq continues to operate as an independent company. It does not describe an acquisition.
Is Cerebras faster than Groq?
On GPT OSS 120B, Artificial Analysis's provider pages showed Cerebras finishing benchmark tasks in about a third of the time on 18 November 2026. Each vendor's own published speeds point the same way. Results change with model, prompt size and load, so test with your own traffic.
Can I still use Llama 3.3 70B on Groq or Cerebras?
Not on a self-serve account. Groq shut it down for free and developer tiers on 16 August 2026 and lists it as enterprise-only. Cerebras deprecated it in February 2026.
Which is cheaper, Groq or Cerebras?
On list prices for GPT OSS 120B, Groq. For Qwen 3.8 27B it depends on your ratio of input to output tokens: Groq is cheaper on input, Cerebras on output.
Does Cerebras have a no-cost tier?
It has a one-time $5 promotional credit that needs a payment method and expires after 30 days. Groq's Free tier is a standing tier with tighter rate limits.
Are the Groq and Cerebras APIs OpenAI-compatible?
Yes. Both document OpenAI-compatible endpoints, so most client libraries work after you change the base URL, API key and model ID. Note the model IDs differ: openai/gpt-oss-120b on Groq and gpt-oss-120b on Cerebras.
Does Groq keep my prompts?
Groq's data page says it does not retain customer data for inference requests by default, with exceptions for features that need stored state, such as batch jobs, and for reliability and abuse investigations.