📰 Key Takeaways

GPT-6 Astra took first place on Retell’s voice agent benchmark, and Retell is also using Codex to generate about $1.6 million in annual revenue run rate per employee. Retell builds AI voice agents, which are used to answer and place phone calls automatically, so the company needs a benchmark to measure how models perform in real-time voice conversations. On this benchmark, GPT-6 Astra scored higher than the other models tested, making it the current top performer. On the other side, Retell uses OpenAI’s coding tool Codex to help with engineering and product development, which lets a small team support a very large revenue scale, with annual revenue run rate averaging $1.6 million per employee. That points to a clear boost in per-person output from AI-assisted development. That said, the original summary only gives these two results. It doesn’t explain the benchmark’s specific tasks, the scoring method, or the actual scores for each model, and it doesn’t say which parts of the development process Codex is used in. So we can’t go into those details here. Please see the original link for the full story.

💬 JudyAI Lab Take

GPT-6 Astra took first place on Retell’s voice agent benchmark, and at the same time Retell is using Codex to generate about $1.6 million in annual revenue run rate per employee. Model results and per-person output showing up together makes this story worth looking at as a pair.

For AI builders, this case points in two directions. First, voice agents have to handle real-time conversations, so Retell measures models with its own benchmark instead of relying only on general-purpose leaderboards. What we can take from this is that picking a model increasingly calls for evaluations that sit close to your actual use case. Second, Retell uses Codex to support engineering and product development, and a small team can carry a very large revenue scale. That shows AI-assisted development is changing the math on per-person output. That said, the original summary doesn’t cover the benchmark’s tasks, the scoring method, or each model’s actual scores, and it doesn’t say which development workflows Codex is used in. So for now these numbers are only a directional reference, and you shouldn’t use them to draw conclusions about your own situation.

My suggestion: pick the 3 most critical use cases for your own product, turn them into a small test set, and use it to compare how different models perform.

📅 Source Info

🔗 Further Reading