
September 2023 Releases
TL;DR: GPT-6.1 Terra is 4th of 50 LLMs tested for voice AI agents on the Voice CX benchmark. It runs at 1,024 ms median time to first token and $1.57 per 1,000 turns, which makes it a fit for verification-heavy, compliance-sensitive flows and a weaker fit for Knowledge Base and tool-heavy resolution.
On October 27, 2026, OpenAI launched GPT-6.1 Terra, a fast, efficient model for detailed visual reasoning, precise computer use, and clear, natural writing. It’s built for tasks like drafting and editing documents, interpreting drawings and screenshots, and completing forms across applications, with standard API input and output token prices at least 50% lower than GPT-5.6 Terra.
But, none of those strengths necessarily translate to strong voice performance. A model that writes a clean email can still stall a live phone call, because voice agents are judged on different things: latency, whether each reply moves the conversation toward the caller's goal, and whether the answer is clear when it's heard rather than read.
That's why we ran GPT-6.1 Terra through the Regal Voice CX Benchmark, which scores models on latency, task progression, clarity, hallucination, and more across real contact center conversation turns. GPT-6.1 Terra currently ranks 4th of 50. Here's the breakdown.

GPT-6.1 Terra posted the best Hallucination sub-score of any model on the benchmark. It was strongest on acknowledgment and verification turns, the moments where an agent confirms a date of birth, reads back a policy number, or repeats what the caller just said.
This is ideal for regulated use cases like insurance enrollment or healthcare scheduling: An agent that never invents a coverage detail or a payment amount is the floor for compliance, and Terra clears it better than any other model we've tested.
Notably, GPT-6.1 Terra has the highest Clarity & Coherence score (88.4) of the top five models, which means it answers efficiently, while speaking in short, natural sentences with clear places to pause.
GPT-6.1 Terra falls short on tool calls and Knowledge Base answers, and that drags down the metric we weigh heavily: Task Progression.
Task Progression measures whether a reply takes the correct, useful next step toward the turn's main goal, without skipping a prerequisite that's needed now or doing something forbidden. In a contact center, this is the difference between a call that resolves and a call that loops.
Compared to GPT-6 Sol, the current leader on our benchmark, GPT-6.1 Terra struggles most with Task Progression on knowledge-answer turns. Here's an example:
.png)
Notably, GPT-6.1 Terra is relatively inexpensive compared to other high-performing models. At $1.57 per 1K turns, GPT-6.1 Terra costs less than a fifth of GPT-5.2 ($8.36) while scoring within 7 points of it.
Terra is a good fit when correctness matters more than autonomy:
Look elsewhere when the agent needs to:
Based on these results, we’re actively evaluating GPT-6.1 Terra on the Regal platform. View the full results here. If you’re curious about our methodology, find more info here.
GPT-6.1 Terra ranks 4th of 50 models on Regal's Voice CX Benchmark. It has the best Hallucination sub-score of any model tested and strong Clarity & Coherence, but trails leaders like GPT-6 Sol on Task Progression, particularly on Knowledge Base answers and tool calls.
GPT-6.1 Terra has a median time to first token of 1,024 ms on Regal's Voice CX Benchmark. That is slower than several top-ranked models, including DeepSeek V4.1 Flash at 353 ms, so it is better suited to verification steps than rapid back-and-forth conversation.
GPT-6.1 Terra costs $1.57 per 1,000 conversation turns on Regal's Voice CX Benchmark. That is less than a fifth of the cost of GPT-5.2 ($8.36), making it one of the stronger value options among top-ranked models.
Task Progression measures whether an AI agent's reply takes the correct, useful next step toward the caller's goal without skipping a required prerequisite or doing something forbidden. It is weighted heavily in Regal's Voice CX Benchmark because it most directly affects whether a call gets resolved.
Yes. In a Regal multi-state agent, each state can use the model best suited to its task, such as a high-precision model like GPT-6.1 Terra for verification and a stronger Task Progression model for resolution.
Ready to see Regal in action?
Book a personalized demo.



