# Nemotron in ElevenAgents: Reasoning Settings and the Real Cost of a Call

The new model is worth evaluating, but a cheaper reasoning engine does not make the entire voice-agent bill cheaper by the same percentage.

By Paras Madan | 2026-11-12

NVIDIA Nemotron 3.5 Super VL is now available in ElevenAgents. ElevenLabs announced the integration on November 11, including an option to disable reasoning for simpler tasks. For teams running support calls, the useful question is whether it reduces the cost of resolving a problem without making the conversation worse.

The [announcement](https://elevenlabs.io/blog/nvidia) reports 70% lower costs than other leading open models on ElevenLabs' customer-support benchmark. It does not identify those comparison models or provide enough billing detail to turn that percentage into a forecast for your account. **Treat it as a reason to run a comparison, not a promised discount on your phone bill.**

## Where to select Nemotron and check its price

The [Models documentation](https://elevenlabs.io/docs/eleven-agents/customization/llm) lists Nemotron 3.5 Super VL as a native option under ElevenLabs. Select it through your agent's model settings rather than treating the announcement as a requirement to deploy your own NVIDIA server.

Before switching production traffic, record your existing model and its settings. Then open **Detailed costs** in the Primary LLM selector. ElevenLabs' [cost help page](https://elevenlabs.io/docs/help-center/product/eleven-agents/how-much-does-eleven-agents-cost) says this shows estimated costs based on the agent's previous behaviour and knowledge base. The public pages reviewed do not establish a Nemotron-specific input/output token tariff, so use the rate shown in your account rather than borrowing a price from another inference provider.

That distinction matters even when the model name is identical. You are evaluating the model through ElevenAgents, with your prompt, knowledge base and conversation history, not purchasing an interchangeable benchmark result.

## Reasoning off is a test condition, not a default recommendation

ElevenLabs documents model-dependent thinking budgets or reasoning-effort settings. A budget of zero disables thinking where supported. Its announcement specifically confirms Nemotron can run without reasoning, but the public guide does not spell out a Nemotron-specific API parameter or complete list of effort levels. Use the controls exposed for the selected model rather than guessing a request field.

Reasoning summaries are a separate debugging option, off by default, and requesting them can add latency. They are not the switch that decides whether the model reasons. Keep that setting consistent when comparing configurations.

For an initial evaluation, I would use one support workflow with two kinds of request: a straightforward invoice lookup and a disputed charge that requires checking policy before taking action. Keep the voice, prompt, tools and test inputs unchanged. Compare the current model against Nemotron with reasoning enabled and disabled. This is a suggested test plan, not a test I have run.

A synthetic case could say: “Invoice A has a $100 subscription and a $20 add-on. The add-on was cancelled after the renewal date. Explain the bill using the supplied cancellation policy.” Success should mean the explanation matches that policy and the agent does not invent a refund. A fluent answer alone is not a pass.

## Calculate total cost, then divide by successful resolutions

ElevenLabs' [current pricing page](https://elevenlabs.io/pricing/agents) separates call minutes from LLM usage. Standard additional minutes cost $0.08 each; burst minutes cost $0.16. External telephone providers bill their own carrier charges separately. These are not Nemotron token prices, and included plan minutes must be accounted for before applying an additional-minute rate.

Here is an illustrative calculation, not a quote or measured result. Suppose 1,000 calls consume three standard additional minutes each, without silence discounts or burst usage. Hosting costs $240. If the model portion were $100 and fell by 70%, the combined hosting-plus-model subtotal would move from $340 to $270. That is about a 21% reduction, before carrier charges and taxes, not 70%.

Now suppose the replacement added 30 seconds to each call. At the same additional-minute rate, those 500 extra minutes would cost $40. The subtotal would become $310, leaving only $30 of the original saving. This does not predict Nemotron's behaviour; it shows why call duration belongs in the model comparison.

Track completed tasks, incorrect actions, transfers, repeated questions and time to the first useful response alongside cost. Divide the full test cost by correctly resolved cases. A model that costs less per token but produces more retries can lose on that measure.

The decision should be narrow: which configuration handles this workflow reliably at the lowest total cost? That is a better reason to switch than a percentage measured on somebody else's support tasks.

I’m Paras. I work at the intersection of AI, engineering and distribution.

I help technical teams turn model choices, product updates and practical engineering lessons into clear stories developers and founders can use.

[Book a 30-minute call](https://cal.com/paras-madan/30min).