Cerebras and Groq offer token pricing, particularly relevant for use with a code broker.
They have some of the highest speeds in the world. Groq and Cerebras offer a token as a service on a limited selection of models. Code agent, voice agent, customer support… There are numerous use cases requiring extremely low latency. While both services offer a similar range of models, the main differences revolve around price and final inference speed. Comparative.
A wider selection of models at Groq
| Models | Brains | Groq |
|---|---|---|
| GPT OSS 120B | x | x |
| GPT OSS 20B | x | |
| GPT OSS Safeguard | x | |
| Kimi K2-0905 1T | x | |
| Llama 3.1 8B | x | x |
| Llama 3.3 70B | x | x |
| Llama 4 Maverick | x | |
| Llama 4 Scout | x | |
| Llama Guard 4 | x | |
| Qwen 3 235B Instruct | x | |
| Qwen 3 235B Thinking | x | |
| Qwen 3 32B | x | x |
| ZAI GLM 4.6 | x |
Groq offers the most extensive selection of models. Among this, GPT-OSS stands out as the open source reference for agentic code applications. The 120B version, accessible via Cerebras and Groq, is to be preferred, even if slightly more expensive (read below). Llama 3.3 70B and Llama 4 Scout offer an attractive balance between performance and cost for general conversational or customer support tasks. The Qwen 3 models, in particular version 235B Thinking and GLM 4.6, are to be used for use cases around reasoning. The selection, whether at Groq or Cerebras, in any case makes it possible to cover a fairly good range of use cases.
Flow: Cerebras wins hands down
| Models | Brain (TPS) | Groq (TPS) |
|---|---|---|
| GPT OSS 120B | 3000 | 500 |
| GPT OSS 20B | 1000 | |
| GPT OSS Safeguard | 1000 | |
| Kimi K2-0905 1T | 200 | |
| Llama 3.1 8B | 2200 | 840 |
| Llama 3.3 70B | 2100 | 394 |
| Llama 4 Maverick | 562 | |
| Llama 4 Scout | 594 | |
| Llama Guard 4 | 325 | |
| Qwen 3 235B Instruct | 1400 | |
| Qwen 3 235B Thinking | 1700 | |
| Qwen 3 32B | 2600 | 662 |
| ZAI GLM 4.6 | NC |
In terms of speed, Cerebras manages to serve models with the highest throughput on the market. GPT OSS 120B, OpenAI’s open source benchmark, is inferred at 3000 tokens per second. A real technical feat that truly saves time in use. The difference with a traditional provider is notable in theory, as in reality. It becomes possible to build applications where speed is a critical variable without any problems. Groq, even if it offers more reasonable speeds, still remains among the fastest providers on the market.
Price: attractive rates at Groq, a monthly offer at Cerebras
Input prices
| Models | Cerebras ($/M tokens) | Groq ($/M tokens) |
|---|---|---|
| GPT OSS 120B | 0.35 | 0.15 |
| GPT OSS 20B | N / A | 0.075 |
| GPT OSS Safeguard 20B | N / A | 0.075 |
| Kimi K2-0905 1T | N / A | 1.00 |
| Llama 3.1 8B | 0.10 | 0.05 |
| Llama 3.3 70B | 0.85 | 0.59 |
| Llama 4 Maverick | N / A | 0.20 |
| Llama 4 Scout | N / A | 0.11 |
| Llama Guard 4 (12B) | N / A | 0.20 |
| Qwen 3 32B | 0.40 | 0.29 |
| Qwen 3 235B Instruct | 0.60 | NP |
| Qwen 3 235B Thinking | 0.60 | NP |
| ZAI GLM 4.6 | 2.25 | NP |
Output prices
| Models | Cerebras ($/M tokens) | Groq ($/M tokens) |
|---|---|---|
| GPT OSS 120B | 0.75 | 0.60 |
| GPT OSS 20B | NP | 0.30 |
| GPT OSS Safeguard 20B | NP | 0.30 |
| Kimi K2-0905 1T | NP | 3.00 |
| Llama 3.1 8B | 0.10 | 0.08 |
| Llama 3.3 70B | 1.20 | 0.79 |
| Llama 4 Maverick | N / A | 0.60 |
| Llama 4 Scout | N / A | 0.34 |
| Llama Guard 4 (12B) | N / A | 0.20 |
| Qwen 3 32B | 0.80 | 0.59 |
| Qwen 3 235B Instruct | 1.20 | NP |
| Qwen 3 235B Thinking | 2.90 | NP |
| ZAI GLM 4.6 | 2.75 | NP |
On paper, Groq displays pricing that is systematically lower than that of Cerebras on all models common to both platforms in both input and output. For GPT-OSS 120B, Groq charges $0.15 per million tokens for input compared to $0.35 at Cerebras, a saving of 57%, while for output, the difference reaches 20% with $0.60 versus $0.75. On Llama 3.3 70B, the difference is 31% in input and 34% in output, a significantly more aggressive positioning on Groq’s side.
However, by calculating the average price-speed ratio on the three shared models (GPT-OSS 120B, Llama 3.1 8B, Llama 3.3 70B), it appears that Cerebras displays an average cost of $0.00017 per token generated per second in output, compared to $0.00135 at Groq. Each token delivered by Cerebras costs almost eight times less at equivalent latency. A difference that radically repositions the competitive advantage in favor of Cerebras for applications where response time is crucial.
Cerebras targets developers, Groq remains more general
Cerebras seems to have perfectly understood the strategic issue that speed represents for developers, particularly in the use of code agents. The company has launched Cerebras Code, a subscription offer specially designed for developers. Two subscriptions are offered: a Pro offer at $50 per month, including one million tokens per minute, 50 requests per minute and a daily quota of 24 million tokens, and a Max offer at $200 per month, offering 1.5 million tokens per minute, 120 requests per minute and 120 million tokens per day.
In practice, during our tests with Cline and Cerebras, we found that speed does indeed change the way you work. Responses arrive almost instantly, which drastically reduces the cycles of iteration, correction and code regeneration. Even if GPT-OSS 120B remains less precise than Claude 4.5 Sonnet or GPT-5 Codex on complex tasks, and therefore generates more errors requiring rereading, the productivity gain remains tangible: we code faster, we explore more, we prototype more.
For its part, Groq adopts a more general approach, addressing a broader spectrum of uses with more moderate pricing. The company highlights a broader range of models, aggressive pricing with performance-cost oriented communication. Groq thus remains relevant for use cases where price is as important a criterion as
