What is SambaNova Cloud?
SambaNova Cloud (also called SambaCloud) is the pay-per-token inference API of SambaNova Systems, which designs its own AI chips, the Reconfigurable Dataflow Unit (RDU). On 7 November 2026 its public model list returned five models: gpt-oss-120b, GLM-5.3, MiniMax-M3.1-Flash, Gemma 4 31B and Llama 3.3 70B. Prices run from $0.22 in and $0.59 out per million tokens for gpt-oss-120b to $2.80 and $8.80 for GLM-5.3.
Artificial Analysis measures gpt-oss-120b on SambaNova at about 417 tokens per second, ahead of Groq's 338 and well behind Cerebras's 1,347. The same hardware is sold for your own data centre as SambaStack, and as SambaManaged for cloud operators who want to resell it.
The company dropped AI training in April 2025, laying off 77 people, about 15% of staff, to focus on inference. Intel discussed buying it for about $1.6 billion in late 2025; instead Intel joined a $350 million round in February 2026 and a multiyear partnership. In July 2026 SambaNova raised $1 billion at an $11 billion valuation.
How much does SambaNova Cloud cost?
| Plan | Price | What you get |
|---|---|---|
| Free | Free | No card linked: 20 requests a minute, 20 a day and 200,000 tokens a day per model |
| gpt-oss-120b | $0.22 in / $0.59 out per 1M tokens | |
| GLM-5.3 | $2.80 in / $8.80 out per 1M tokens | $0.14 per 1M cached input tokens |
| MiniMax-M3.1-Flash | $0.30 in / $1.20 out per 1M tokens | Preview model, 1M-token context; $0.03 per 1M cached input tokens |
| Enterprise | Quote | Subscription with production rate limits, on-request models and bring-your-own-checkpoint |
Pay-as-you-go per 1M tokens: gpt-oss-120b $0.22 in / $0.59 out; Gemma 4 31B $0.38 / $1.15; MiniMax-M3.1-Flash $0.30 / $1.20 ($0.03 cached input); Llama 3.3 70B $0.60 / $1.20; GLM-5.3 $2.80 / $8.80 ($0.14 cached input). A free tier without a card allows 20 requests a day per model. Enterprise subscriptions by quote.
Is SambaNova Cloud worth it?
If you are choosing where to run an open model, the speed charts make it look like a two-horse race between Groq and Cerebras. There is a third chipmaker selling tokens, and on some models it is quicker than Groq.
SambaNova Cloud runs open models on SambaNova's own chips, which keep a whole model close to the processor and can swap between several models quickly. Artificial Analysis measures its gpt-oss-120b at about 417 tokens per second, against 338 for Groq and 65 for Together AI, though Cerebras reaches 1,347. At $0.22 per million input tokens and $0.59 output, it costs less than Cerebras's $0.35 and $0.75 and a little more than Groq's $0.15 on input. Artificial Analysis also scores its gpt-oss-120b answers at 98.5% of the reference model's accuracy, the highest of the fast providers.
The menu is short and moves. The live model list on 7 November 2026 had five models, including GLM-5.3 at $2.80 and $8.80 and a preview MiniMax-M3.1-Flash with a 1-million-token context, while the documentation still listed DeepSeek models the API no longer returned. Preview models, SambaNova says, "may be removed at short notice", and Llama 3.3 70B caps each answer at 3,072 tokens.
The catch: self-serve capacity is small. With a card linked, the Developer tier allows 60 requests a minute and 12,000 a day on most models, and 20 million tokens a day across all of them. Anything bigger means an Enterprise subscription with a quoted price. On data, SambaNova's product page says it never collects your prompts, but its licence agreement gives no retention period and lets it use query logs to improve its service. We found no public SOC 2 report, so get the promise in a contract.
Our take: SambaNova is the vendor to call when you want fast inference on hardware you own. SambaStack runs the same chips in your data centre, air-gapped if needed, and JPMorgan Chase is buying it on those terms. As a self-serve API it is a sensible second route for gpt-oss, not the place to put your whole product.
Where it earns its place
A five-person team building a support chatbot on gpt-oss-120b
Add a fast fallback provider without paying Cerebras prices
Point the OpenAI-compatible client at SambaNova when the main provider returns errors. At $0.22 and $0.59 per million tokens, 100 million input and 20 million output tokens a month cost about $34, well inside the 20-million-a-day Developer cap.
An infrastructure lead at a bank or public agency
Run fast open-model inference where no prompt leaves the building
SambaStack installs SambaNova's racks on premises, with air-gapped deployment since its March 2026 release; JPMorgan Chase is deploying SN40L and SN50 systems this way. Pricing is by quote.
A developer building a long-document agent
Read very long inputs quickly and cheaply
MiniMax-M3.1-Flash accepts up to 1 million tokens at $0.30 per million input, and $0.03 when the prefix is cached. It is a preview model, so keep a second provider ready.
Before you commit
- Setup
- Create an account and an API key, then set the base URL to api.sambanova.ai/v1 in any OpenAI SDK; linking a card moves you from the free tier to the Developer tier.
- Integrations
- OpenAI-compatible chat API, Anthropic Messages API since July 2026, a SambaNova SDK, OpenRouter and AWS PrivateLink.
- Data and privacy
- SambaNova says it never collects your prompts; its licence lets it store content only to provide the service but gives no retention period, and it may use query logs to improve the service. We found no public SOC 2 report or DPA.
- Cost at scale
- 2 billion input and 500 million output tokens a month on gpt-oss-120b would cost about $735, against about $1,075 on Cerebras and $600 on Groq, but that volume is over the Developer tier's 20-million-a-day cap and needs an Enterprise subscription.
- Limits
- Developer tier: 60 requests a minute and 12,000 a day per model (240 and 48,000 on Llama 3.3 70B), 20 million tokens a day in total. Context 131,072 tokens on most models, 1 million on MiniMax-M3.1-Flash.
What’s good
- About 417 tokens/s on gpt-oss-120b, per Artificial Analysis
- gpt-oss-120b at $0.22 in and $0.59 out per million tokens
- Highest gpt-oss-120b accuracy score of the fast providers on Artificial Analysis
- Same chips available on premises through SambaStack
What’s not
- Developer tier capped at 20 million tokens a day
- Five models, and the docs disagree with the live list
- No retention period in the licence; no public SOC 2 report
- Llama 3.3 70B answers capped at 3,072 tokens
Best for
Teams that want fast gpt-oss or GLM inference as a second provider, or fast open-model inference on hardware in their own data centre.
SambaNova Cloud alternatives
SambaNova Cloud: common questions
- Is SambaNova Cloud worth it?
- SambaNova sells fast gpt-oss tokens for less than Cerebras, but the self-serve limits are small enough that any real workload ends in a sales call. We rate it 3.8 out of 5.
- How much does SambaNova Cloud cost?
- Pay-as-you-go per 1M tokens: gpt-oss-120b $0.22 in / $0.59 out; Gemma 4 31B $0.38 / $1.15; MiniMax-M3.1-Flash $0.30 / $1.20 ($0.03 cached input); Llama 3.3 70B $0.60 / $1.20; GLM-5.3 $2.80 / $8.80 ($0.14 cached input). A free tier without a card allows 20 requests a day per model. Enterprise subscriptions by quote.
- What are the downsides of SambaNova Cloud?
- Developer tier capped at 20 million tokens a day; five models, and the docs disagree with the live list; no retention period in the licence; no public SOC 2 report; llama 3.3 70B answers capped at 3,072 tokens.
- Who is SambaNova Cloud best for?
- Teams that want fast gpt-oss or GLM inference as a second provider, or fast open-model inference on hardware in their own data centre.
- How much does SambaNova Cloud cost for a team or in production?
- 2 billion input and 500 million output tokens a month on gpt-oss-120b would cost about $735, against about $1,075 on Cerebras and $600 on Groq, but that volume is over the Developer tier's 20-million-a-day cap and needs an Enterprise subscription.
- Is SambaNova Cloud safe for company data?
- SambaNova says it never collects your prompts; its licence lets it store content only to provide the service but gives no retention period, and it may use query logs to improve the service. We found no public SOC 2 report or DPA.
- What are the best SambaNova Cloud alternatives?
- The closest alternatives we have reviewed are Cerebras, Groq, Together AI and Fireworks AI.
Answers checked Nov 7, 2026.