Best LLMs for Real-Time Voice AI in 2025: Vendor Survey Results

Real-Time Voice AI in 2025: Which LLMs Top the Vendor Short-List?

When you need to know which large-language model (LLM) actually delivers in real-time Voice AI, ask the people who build and run these systems every day. That's exactly what I did. I polled a group of conversational-AI vendors and asked a single question: "Which LLM will you choose for production Voice AI in 2025?"

The survey results

1. GPT-4o — 3 votes
2. GPT-4o Mini — 3 votes
3. Gemini 2.0 Flash — 2 votes
4. GPT-4.1 — 1 vote
5. Own in-house model — 1 vote

Nine out of ten selections went to commercial, state-of-the-art models—clear proof that cutting-edge capabilities remain non-negotiable for high-quality voice interactions.

Four key takeaways

1. State-of-the-art models are still mandatory

The best currently available models are essential for natural, reliable conversations. Vendors overwhelmingly prefer frontier LLMs over older or fully open-source alternatives.

2. Success is a balancing act of speed, cost, and capabilities

Real-time Voice AI must juggle low latency, solid instruction following, dependable tool calling, and reasonable cost. All three leading models check these boxes, but each leans on a different strength:

• Latency: GPT-4o Mini responds the fastest.
• Pricing: Gemini 2.0 Flash is the most affordable, closely followed by GPT-4o Mini and well ahead of GPT-4o.
• Instruction following & tool calls: Gemini 2.0 Flash performs on par with GPT-4o, while GPT-4o Mini trails slightly.

The "right" choice hinges on your target use case, budget, and which providers you can access.

3. Keep an eye on GPT-4.1—especially the Mini variant

Although the GPT-4.1 family was built with developers in mind, GPT-4.1 Mini offers an intriguing mix of low latency, strong instruction handling, and competitive pricing. It hasn't reached the leaderboard yet, and I haven't run it in production, but initial benchmarks look promising.

I've been recommended to also keep an eye on Amazon Nova Pro and Gemini Flash 2.5, which is currently in preview.

4. Speech-to-speech models aren't production-ready—yet

End-to-end speech-to-speech (STS) models remain in beta, wrestling with early bugs and premium pricing. Long-term, they could upend the entire Voice-AI stack, but for now they stay on the sidelines.