Not tested by us yet. Tell-1 is a model inside Retell's voice-agent platform, with no public API or model card that our harness could call. Everything below comes from Retell's benchmark changelog entry, its GPT-Live-1 changelog entry and its vCX benchmark site. We'll update this post when Tell-1 can be called directly or Retell publishes the benchmark data.
Retell AI, a company that sells phone voice agents, says its own model, Tell-1, ranks first overall on the Retell vCX Benchmark as of November 2026. Per Retell, Tell-1 leads on colloquial speech and latency, while other models still lead on grounding, instruction following and tool calling. If you run contact-center voice agents on Retell, Tell-1 is worth a trial on your own calls. If you need numbers you can verify, there are none yet.
| Tell-1 | |
|---|---|
| Maker | Retell AI |
| Type | Voice model for live phone calls |
| Announced | 2026-11-12 (press coverage); benchmark note says "as of November 2026" |
| Claimed strengths | Colloquial speech, latency |
| Claimed weaknesses | Grounding, instruction following, tool calling (other models lead) |
| Price | Not published |
| Access | Retell platform; terms not stated |
| Independent results | None found |
What Retell actually claims
The claim sits in an update to a September 29 changelog entry about a public voice AI model benchmark. It calls Tell-1 "Retell's own model" and says it ranks #1 overall. The benchmark page says it covers instruction following, grounding, colloquialism and latency, and that results are updated as new models ship.
Read that carefully. Retell makes the model, runs the benchmark and decides the scoring. The pages we could read give no scores, no sample size, no scoring method and no list of competing models. The benchmark site describes comparing "quality, latency, and reliability across top providers", but the leaderboard did not load as text for us, so we could not check the figures.
The part that matters for production
The admission is more useful than the headline. Retell says other models still lead on grounding, instruction following and tool calling. In a contact center those three decide whether the agent reads the right policy, follows your script and books the appointment. Colloquial tone and speed decide whether the call feels natural. A model that wins overall by being fast and chatty but trails on tool calling is a trade, not a free upgrade.
This is also why Retell's design keeps two layers. Its GPT-Live-1 entry describes a full-duplex voice model that talks to the caller directly, while a text LLM of your choice runs the prompt, tools and knowledge base. It works on single-prompt voice agents on paid accounts, not on conversation flows, multi-prompt, transfer or chat agents. Retell has not said whether Tell-1 uses the same split or its limits.
How it compares on paper
We can compare only what is published, and little is. Tell-1 and GPT-Live-1 both live on Retell. GPT-Live-1 is documented as a voice layer over your own text LLM, and Tell-1 is described only through benchmark rankings. Neither entry states a price. Among other new voice releases we have covered, the closest read is Chariot-1 TTS, which is a speech generator rather than a full call agent, so it answers a different question.
Who should care
- Retell customers: run Tell-1 against your current setup on 20 to 30 real recorded calls, and score tool-call success separately from tone.
- Teams comparing voice platforms: treat a vendor-run leaderboard as a lead, not evidence. Ask for the call set, scoring rubric and raw latency figures.
- Everyone else: nothing to act on until there is a price, an API or independent numbers.
What would change this post
We will re-test when Tell-1 is reachable outside Retell's platform, or when the vCX data and method are published in full. Until then, every performance statement here is Retell's, unverified by us. For another model category judged mostly on vendor claims, see our Kanana-2-v-decide teardown.