Introducing Tell-1, Retell’s LLM purpose-built for contact centers.
@retellai4K followers
72,242 views
26 Comments
Comments from the original post on X (Twitter)

@retellai
Tell-1 by the numbers: → 94.4% average score on Retell’s vCX benchmark (#1 overall) → ~150ms time to first token → $0.035 per minute Explore the full results and methodology:
10 likes

@retellai
Tell-1 has also been evaluated by independent testing firms including @cekuraAi , @covaldev, @getbluejay_ai and @HammingAI . Explore their results: → Cekura: → Coval: → Bluejay: → Hamming AI:
12 likes

@retellai
What happens while Tell-1 is speaking? As its first sentence plays, it reasons about which business rule applies, what information is missing, and whether to call a tool. Go deeper into the training, inference, and evaluation behind Tell-1:
8 likes

@retellai
Tell-1 is live on Retell. Try it on your own calls - with your business rules, workflows, and the unexpected turns your customers bring. We’d love to hear what you think.
6 likes

@KavakErkam
@zexiazh Is it going to be available as API?
1 like

@brutallee666
@zexiazh This model is blazingly fast
1 like

@shanaygaitonde
@zexiazh what happens when the reasoning catches that the first answer was wrong
1 like

@vitobotta
@zexiazh Answering first and reasoning while it talks is a clever latency trick. What happens when it realises it's wrong mid-sentence?
1 like

@rarestomaa
@zexiazh what happens when the reasoning disagrees with what it already said out loud?
1 like

@I_Aditya_
@zexiazh Excited to give it a shot 🚀

@CholeSyntax
@zexiazh Answer first reason while speaking that's exactly how real contact centers should feel

@YXxquf
@zexiazh Reasoning while speaking makes interruptions especially interesting: the customer can change the request before the thought finishes. Does your benchmark test whether Tell-1 cancels an outdated tool action after a mid-sentence correction?

@Abdulrahman_NK_
@zexiazh Answering first and reasoning while speaking is a smart fix for voice latency. In a contact center a 2 second pause feels like forever.

@Dinamo51648539
@zexiazh Your technical note says the lead bonus trades latency for accuracy. I'd score the opening sentence separately: premature promises on refunds, bookings or identity checks before the rule or tool result arrives. A correction in sentence two doesn't undo what the caller heard.

@eStomes
@zexiazh Very impressive initial tests. Congrats.

@Grumpysmurf_xyz
@zexiazh exactly how skilled operators talk

@tfmlab
@zexiazh Answer-first, reason-while-speaking commits spoken words before reasoning finishes. Worth reporting how often later reasoning contradicts what was already said, and how the agent repairs it mid-turn, next to time-to-first-audio. Latency gains matter less if corrections rise.

@EllaTech0
@zexiazh This is so true smart speed changes contact centers completely

@ShivamGuptaim
@zexiazh The first sentence needs to stay honest while the reasoning catches up. “Let me check whether that qualifies” buys time; “yes, we can refund that” creates a promise. I’d test premature commitments explicitly, not just whether the eventual tool call was correct.

@BaranwithAi
@zexiazh Answering first then reasoning while speaking is clever design





