Tell-1: the First LLM for Contact Centers
Ranking #1 on cost, intelligence, and latency for contact centers, thinking while it speaks.
Watch the interview · 3 min Model Architecture
Interleaved reasoning
Most voice agents pause to think. Over a whole call, that sounds robotic.
Tell-1 answers right away and reasons while it speaks: which rule applies, what is missing, when to call a tool.
Read the technical blog →Evals
Ahead of every model on the board
Top five on all four skills in the Retell vCX benchmark. First for sounding natural.
Where Tell-1 ranks
Tell-1 Other modelsAverage score
94.4%Across all four skills, the top average on the board.
- Top 5in every skill
- #1in colloquial speech
Highest average score, fastest first sentence
Average Retell vCX score against median time to first sentence (P50, ms)
The score averages tool calling, hallucination, instruction following and colloquial accuracy. Source: bench.retellai.com, read Nov 12, 2026.
Third-party Benchmarks
Beyond our own benchmark
Results from independent voice AI testing teams.
Tell-1 on Sierra's τ²-bench and other public benchmarks
Score out of 100, same greedy decoding on every run. None were used in training.
- τ²-bench airline & retail
- Serving a simulated customer with tools and a policy to follow
- Multi-IF
- Following instructions over three turns in eight languages
- GPQA Diamond
- Graduate-level science questions written to resist web search
Cekura Retell, running Tell-1, ranked first of 10 voice platforms for reliability.
See Cekura's benchmark →
Bluejay Ranks the LLMs behind phone agents on healthcare, legal and support calls.
See Bluejay's benchmark →
Hamming Grades the LLMs behind voice agents on response time, price, logic, phrasing and security.
See Hamming's benchmark →At Scale
Production Performance
Follows the right playbook
A contact center can have a thousand intents, each with its own SOP.
Tell-1 identifies the right SOP, follows it step by step and brings the caller back when they drift.
Fast, smart, affordable
No trade-off between latency, intelligence and cost.
Tell-1 runs next to your voice agent on Retell's own inference stack, averaged 94.4% on the Retell vCX benchmark, and is built to keep latency and cost steady at a million calls a month.