| Model | Best for | Voice CX Score | Time to first token | Cost per 1K turns |
|---|
| GPT-6 Sol | Best overall | 786 | 870 ms | $19.88 |
| DeepSeek V4.1 Flash | Best value and fastest response | 783 | 353 ms | $1.98 |
| GPT-5.2 | Tool calling and multi-state agents | 779 | 904 ms | $8.36 |
| GPT-6.1 Terra | Accuracy and fewest hallucinations | 772 | 1,024 ms | $1.57 |
| GLM 5.3 Flash | Open-source | 767 | 984 ms | $1.34 |
| GPT-5.6 Luna | High-volume production calls | 766 | 988 ms | $1.18 |
| Claude Haiku 5.5 | Teams standardized on Claude | 746 | 826 ms | $2.54 |
How We Ranked These Models
Every ranking in this article comes from the Regal Voice CX Benchmark, which tests how LLMs perform inside live voice conversations. Nearly 50 models from OpenAI, Google, Anthropic, DeepSeek, Zhipu AI and others run through the same inbound and outbound call scenarios, built mostly from real call transcripts. Scenarios cover customer support and lead qualification, on both single-state and multi-state agents. New models are added as they ship.
Every turn is scored on five dimensions:
- Task progression: does the turn move the call toward its goal?
- Latency: median time to first token, meaning how long the caller waits before the agent starts responding.
- Accuracy and safety: is the agent correct, compliant, careful with sensitive data, and reliable with tools?
- Clarity and coherence: is the response easy to follow when it's spoken aloud?
- Personalization and engagement: does the agent adapt to the caller and keep the call moving?
These combine into a single Voice CX Score out of 1,000, weighted most heavily toward task progression and latency, because those two best predict whether a model holds up in production.
How each turn is graded: before scoring, every turn is classified by what it's trying to do, such as a greeting, identity verification, a knowledge answer, objection handling or appointment scheduling. The type decides which metrics apply. For example, hallucination is checked on knowledge answers and verification rigor on identity checks, so a greeting is never marked down for failing a knowledge check.
Why latency matters so much: Claude Opus 5, Gemini 3.1 Pro and Claude Sonnet 4.6 all score close to or above GPT-6 Sol on task progression, but they take 1.9 to 5.4 seconds to start responding and score 623 to 633. On a phone call, two seconds of silence sounds like a dropped line. For the engineering side of this, see how Regal cut AI agent latency by 26%.
How to Choose an LLM for Your Voice AI Agent
- Most production agents: start with GPT-5.6 Luna. It's Regal's default model, scores in the top 10, and costs $1.18 per 1,000 turns.
- Agents that call many actions or move through multiple states: GPT-5.2 is the most reliable at calling tools when it should.
- Knowledge-heavy or regulated calls, where a made-up answer is the biggest risk: GPT-6.1 Terra has the lowest hallucination score of any model tested.
- Tight budget or a hard latency limit: DeepSeek V4.1 Flash is the fastest model tested and scores within 4 points of the leader.
- Open weights or self-hosting: GLM 5.3 Flash is the highest-scoring open-source model.
- Whatever you pick: retest your prompts and run simulations before switching models. A prompt tuned for one model rarely transfers cleanly to another.
Not every model here is available on every platform. For the models you can select in Regal today, see the Regal LLM models guide.
1. GPT-6 Sol — Best Overall LLM for Voice AI
GPT-6 Sol has the highest Voice CX Score in the benchmark, 786 out of 1,000, with a median time to first token of 870 ms at reasoning set to none. It ranks first on voice format compliance, meaning its answers are written to be spoken rather than read, and second on guardrail compliance. That makes it a strong pick for regulated conversations in insurance, lending and healthcare.
The tradeoff: at $19.88 per 1,000 conversation turns, it costs about 10x more than DeepSeek V4.1 Flash, which scores only 3 points lower.
2. DeepSeek V4.1 Flash — Best Value and Fastest LLM for Voice AI
DeepSeek V4.1 Flash scores 783, the highest of any lightweight model and within 4 points of the leader, at $1.98 per 1,000 turns. It's also the fastest model in the benchmark, starting to respond in 353 ms. Its task progression (0.614) is among the best tested, and it's good at holding back tool calls when they aren't needed.
The tradeoff: it calls required tools unevenly, especially in multi-state flows, and its phrasing can sound more scripted than the leaders'. A small share of its answers ran to the output limit with empty padding, so test it on your own call flows first.
3. GPT-5.2 — Best for Tool Calling and Multi-State Agents
GPT-5.2 scores 779 and ranks first in the benchmark at calling tools when a turn requires one (0.858). It's also strong at step sequencing and moving multi-state conversations forward, with task progression of 0.748 on multi-state flows against an average of 0.615. That suits agents that book appointments, update records or hand off between states, at $8.36 per 1,000 turns and 904 ms to first token.
The tradeoff: it's quick to wrap up. It ranks near the bottom on premature closure, so spell out in the prompt when the agent may end the call.
4. GPT-6.1 Terra — Best for Accuracy
GPT-6.1 Terra has the lowest hallucination score of any model tested and ranks second on response efficiency, with a Voice CX Score of 772. It's especially strong on acknowledgment and verification turns. At $1.57 per 1,000 turns, it scores 16 points higher than GPT-5.6 Terra for 86% less.
The tradeoff: it misses more required tool calls than the leaders, especially on appointment-scheduling turns. On knowledge-answer turns it can hold back information it has been given, so test those flows before going live.
5. GLM 5.3 Flash — Best Open-Source LLM for Voice AI
GLM 5.3 Flash is the highest-scoring open-source model at 767, ahead of GLM 4.7 (732) and every other open-weight model tested, at $1.34 per 1,000 turns. It has one of the highest answer-completeness scores in the benchmark and ranks near the top on response efficiency, voice format compliance and calling tools when it should.
The tradeoff: it calls tools when it shouldn't more often than most models, especially on simple acknowledgment turns, and its 984 ms response time is near the upper limit callers tolerate. Self-hosting or running it through an inference provider also adds operational overhead.
6. GPT-5.6 Luna — Best for High-Volume Production Calls
GPT-5.6 Luna scores 766 at $1.18 per 1,000 turns, the lowest cost of any model in the top 10. It's the default model in the Regal Agent Builder, and it's especially strong on compliance-script turns (0.882 task progression against an average of 0.689) and verification turns. That fits high-volume programs like qualification, scheduling, confirmations and payment reminders.
The tradeoff: it's weaker on objection handling and on complete answers to open-ended questions. For long, multi-step calls where finishing the task matters most, GPT-5.2 or GPT-6 Sol are stronger.
7. Claude Haiku 5.5 — Best Claude Model for Voice AI
Claude Haiku 5.5 scores 746, 35 points higher than Claude Haiku 4.5, with an 826 ms response time and a cost of $2.54 per 1,000 turns. It's among the best tested at understanding caller intent and keeping context across a call. Larger Claude models are held back by speed: Claude Sonnet 4.6 scores 623 because it takes 2.2 seconds to start responding.
The tradeoff: its task progression (0.525) trails the leaders, and it calls tools when it shouldn't more often than most. Instruct it to confirm with the caller before any action that changes data.
Schedule a demo with Regal