Groq and Cerebras Both Pulled Llama 3.3 70B Off Their Price Lists

21 October 2026 · Site Methodology & Corrections · 5 min read

I was updating our Groq pricing page and went to re-check the Llama 3.3 70B rate I'd verified back in August. It's gone. Groq's own model table now lists llama-3.3-70b-versatile as "Enterprise," with both the price and the rate limit columns reading "Contact Sales." I checked Cerebras next, since it's the other provider built entirely around fast open-model inference. Same story: Llama 3.3 70B is still technically supported, but only through Dedicated Inference -- the Enterprise tier, priced "Contact us for a quote." Neither company's self-serve, pay-as-you-go tier publishes a Llama 70B price anymore.

What used to be public, and what replaced it

In September I wrote about Groq's Llama 70B pricing at $0,59 de entrada / $0,79 de salida por 1 millón de tokens. That number isn't wrong because I mis-measured it -- it's wrong because Groq removed it. Pulling the current GroqCloud model table, the only two text models with a public per-token price are GPT-OSS-120B and GPT-OSS-20B:

Mixtral 8x7B, which used to round out Groq's self-serve lineup, isn't in the current model list at all -- featured, production or preview tiers. Cerebras tells the same story from the other side: its public Developer Tier pricing table lists exactly two models with a price -- GPT-OSS-120B at $0.35/$0.75 per 1M, and Qwen3.8-27B at $0.99/$1.49 per 1M. Llama 3.3 70B appears only in Cerebras's Dedicated Inference supported-models list, which is the "contact us for a quote" product.

Where Llama 70B is still a published price

It hasn't vanished from the market -- just from the two speed-focused providers. SambaNova and Together AI both still quote it in the open, at noticeably different rates:

If Llama 3.3 70B specifically is what your app needs, SambaNova is the cheaper of the two public options by a wide margin -- about 42% less than Together on input, and close on output once you weigh a realistic 3:1 input-heavy workload.

The model everyone actually prices now: GPT-OSS-120B

Since GPT-OSS-120B is the one ~120B-class open model all four providers still publish a self-serve rate for, it's become the real apples-to-apples comparison -- not by my choice, but because it's the only one left standing on every price list:

Groq and Together land on an identical headline rate. SambaNova is close behind and actually cheaper on output. Cerebras is the most expensive of the four on paper -- but it's also the only one advertising a specific speed number on the same page, and 3,000 tokens/second is roughly 6x Groq's published 500. Whether that's worth the premium depends entirely on whether your workload is latency-bound or cost-bound; a voice agent racing to finish a sentence before the user notices a pause cares about the first number, a batch summarization job cares about the second.

Lo que hice al respecto

I rewrote the Groq pricing page: the table now shows GPT-OSS-120B and GPT-OSS-20B as the self-serve models, with an explicit note that Llama moved to Enterprise/Contact Sales, and I swapped the quick-start curl example off llama-3.1-8b-instant since that model ID no longer works on a self-serve key. I didn't touch the September Llama 70B comparison post itself -- it was accurate when published -- but added a note pointing here, since its headline Groq number is the one that's since disappeared.

Misma lección que la DeepSeek Search/Fetch pricing piece last week: a provider's price list is a snapshot, not a contract, and the fast-inference market in particular is still rearranging what it charges for which model every few weeks.

• Groq API pricing guide — updated with the current self-serve model list.
• Cerebras API pricing guide — GPT-OSS-120B and Qwen3.8-27B, the two models Cerebras still prices publicly.
• Guía de precios de Together AI — still lists Llama 3.3 70B at a public rate.
• Calculadora de costos de API de IA — run your own volume against GPT-OSS-120B, Llama 70B or any model we track.

¿Es nuevo en los precios de API basados en el uso en general? Comience con el guías gratuitas de costos de API.

Despliéguelo usted mismo: DigitalOcean — $200 de crédito gratis ↗ · VPS Hostinger ↗

Checked directly against Groq's console.groq.com/docs/models table, Cerebras's cerebras.ai/pricing and inference-docs.cerebras.ai dedicated-tier model list, SambaNova's cloud.sambanova.ai/pricing, and Together AI's together.ai/pricing, 2026-10-21. Reference figures, not live quotes -- confirm current terms directly with each provider before budgeting.