Retell on November 12, 2026 launched Tell-1, a language model built for live contact-center phone calls rather than chat windows. The company describes it as trained on hundreds of millions of calls, tuned for partial answers, detours, and mid-call tools, with Tell 1 Beta priced at $0.035 per minute on its pricing page. The same announcement introduced Retell Dialer, an outbound campaign orchestrator for voice and SMS agents.
For builders shipping phone agents, the news is a bet that telephony latency and turn-taking beat raw benchmark scores from general-purpose chat models.
TL;DR #
| Item | Retell says (Nov 12, 2026) |
|---|---|
| Model | Tell-1 voice LM for live calls | | Architecture pitch | Interleaved reasoning — thinks while speaking | | Latency claim | Under 100 ms on live bench;under 300 ms at peak | | Price | $0.035/min Tell 1 Beta voice;$0.011 per chat AI message | | Benchmark | #1 overall on RetellvCX board (vendor-owned) | | Companion ship | Retell Dialer for outbound campaigns |
Why contact centers need a different model shape #
Chat-optimized models excel at long context and markdown. A PSTN caller punishes anything that sounds like it is reading an essay. Common failure modes Retell names in its Business Wire release include:
- Half answers — caller gives only a first name; the agent must ask for the rest without resetting the form.
- Detours — off-topic questions that need acknowledgement before returning to the task.
- Tool timing — lookup or booking actions once enough slots are filled, without dead air.
Those are the same stresses covered in explainx.ai guides on Confucius4 streaming for voice agents and GPT-Live-style duplex models, but Tell-1 is Retell’s attempt to own the default brain instead of marking up OpenAI or Anthropic per-minute fees.
Interleaved reasoning on the wire #
Retell’s headline mechanism is interleaved reasoning: the model updates its internal plan while audio is already flowing, rather than waiting for a full completion token batch.
In product terms that means policy exceptions, missing fields, and tool eligibility are tracked across minutes, not isolated turns. That mirrors how human reps juggle scripts — and how decision models gate cheaper classifiers upstream of expensive speech stacks.
Benchmarks: strong vendor story, independent gap #
Retell maintains a public vCX benchmark comparing dozens of hosted model configurations. Its November changelog states Tell-1 ranks #1 overall, leading colloquial speech and latency, while other models still win some columns on grounding or tool calling.
The best LLM for voice agents post, updated November 11, 2026, claims Tell-1 scores 94.45% averaged across instruction following, grounding, tool calling, and colloquial quality, with lowest median latency among 62 listed configs. Those numbers are Retell-measured, not third-party.
Compare with adjacent launches:
- [ElevenLabs Scribe v3](https://explainx.ai/blog/elevenlabs-scribe-v3-v3-live-speech-to-text-turn-detection-2026) attacks**ASR and turn detection** , not dialog policy.
- [Fish Audio S2.1 Pro](https://explainx.ai/blog/fish-audio-52m-seed-s2-1-pro-july-2026) chases**TTS quality** for studios.
- Tell-1 sits in thedialog + policy layer Retell already bills per minute.
Retell Tell-1 voice model latency claims for live contact-center phone agents
Pricing and when Tell-1 beats frontier per-minute models #
On the pricing table, Tell 1 Beta sits at $0.035/minute, below many listed frontier entries such as GPT-6 Astra at $0.32/minute or Claude 5.5 Opus at $0.128/minute. Totals still stack with voice engine, TTS, telephony, and add-ons like PII removal or safety guardrails.
For teams already on Retell, switching default LLM to Tell-1 is a configuration change, not a migration. For teams comparing platforms, run your own labeled call set — vendor boards overweight colloquial speech and latency weights Retell chooses.
Retell Dialer: outbound orchestration #
Retell Dialer, launched the same day, connects outbound lists to Tell-1-driven conversations and downstream SMS nurture agents. Retell emphasizes:
- Callbacks at callee-local time instead of blind retries.
- Qualified handoffs to human or specialist agents.
- CRM hooks with rotation, local presence, and spam monitoring.
That pairs with AWS and telephony trends we track in Bedrock agent tooling, but Dialer is Retell-native campaign glue rather than a generic LLM API.
How to try Tell-1 this week #
- Open an existing single-prompt voice agent on a paid Retell account (flows and multi-prompt agents may still require frontier models — check Retell docs for your agent type).
- Select Tell 1 Beta in the LLM dropdown and keep your prior model as ashadow for A/B comparison.
- Log median time-to-first-audio-token ,tool success rate , andhuman transfer rate over at least200 calls before switching production traffic.
- If you rely on OpenRouter-style model fan-out , remember Tell-1 isnot an exportable weight — budget stays inside Retell meters.
Retell Tell-1 pricing on the voice agent stack versus frontier models per minute
vCX benchmark dimensions (how Retell scores voice) #
Retell’s public benchmark page breaks “good on a phone call” into columns that chat leaderboards rarely test:
| Dimension | Why it matters on PSTN |
|---|---|
| Instruction following | Sticking to disclosure scripts regulators require |
| Grounding | Citing account facts without inventing balances |
| Tool calling | CRM updates, calendar holds, payment APIs |
| Colloquial quality | Sounding human on low-bitrate audio |
| Latency | Holding barge-in and backchannel timing |
Tell-1’s claimed #1 overall rank is a weighted blend of those columns on Retell infrastructure. If you already run GPT-6 Luna or Gemini 3.5 Flash Lite on Retell for cost reasons, re-score on your compliance script before switching — grounding weights differ by industry.
Competitive landscape in November 2026 #
Voice agents sit between CRM giants and model labs:
- [Salesforce’s Cresta acquisition path](https://explainx.ai/blog/salesforce-cresta-acquisition-voice-agents-service-cloud-2-7-billion-reported-2026) pushes**voice-first service cloud** consolidation.
- [ElevenLabs Scribe v3](https://explainx.ai/blog/elevenlabs-scribe-v3-v3-live-speech-to-text-turn-detection-2026) improves**ears** (ASR + turn detection).
- Tell-1 tries to own thebrain on a telephony-native host.
Retell’s bet is vertical integration: LLM + telephony + orchestration in one bill, similar to how AWS Bedrock AgentCore bundles runtime + governance for text agents.
Dialer workflows builders can copy #
Retell Dialer’s examples in the launch materials map cleanly to agent design patterns:
- Requested callback — parse local time from Tell-1 transcript, write CRM task, suppress duplicate dials.
- Qualified lead — transfer to human or high-tier agent withcontext bundle (Retell conversation id).
- Nurture path — trigger SMS agent withshorter Tell-1 or chat pricing ($0.011/msg).
If you orchestrate multi-agent flows, mirror that split: cheap text decision model for SMS, Tell-1 for voice, echoing decision-model routing guides.
Stack economics: voice engine + LLM + telephony #
Retell invoices in layers. Tell-1’s $0.035/min is only the LLM row. Production calls also incur:
- Retell Voice Infra at**$0.055/min** on the pricing page
- TTS choices from**$0.015** to**$0.10/min** depending on ElevenLabs tier
- Telephony around**$0.015/min** for U.S. Twilio routes
- Optional PII removal ,safety guardrails , andQA surcharges
A back-of-envelope minute can exceed $0.12 before carrier taxes. That is still often below stacking GPT-6 Astra voice pricing on the same table, but not automatically cheaper than Gemini 3.5 Flash Lite at $0.0096/min LLM if your script tolerates weaker grounding.
Finance teams should model cost per completed task, not cost per minute — the same lesson from Bedrock cost-per-task baselines.
Compliance and enterprise knobs #
Retell’s enterprise plan advertises SSO, RBAC, custom MSA/DPA, and high concurrent call caps. For regulated callers (health, finance), pair Tell-1 with:
- Recorded consent flows required in your jurisdiction
- Knowledge base add-on (+$0.005/min) when answers must cite uploaded policy PDFs
- Human transfer targets when confidence drops — measure how often Tell-1 requests handoff versus frontier models
Those operational metrics usually matter more to legal review than a 94.45% vendor benchmark average.
Build checklist before production cutover #
- Record baseline median latency andtransfer rate on your incumbent LLM for one week.
- Enable Tell-1 on 10% traffic withrecording consent flags unchanged.
- Compare tool error rate on your top three API tools (payments, account lookup, scheduling).
- Reconcile per-minute invoice including voice engine + TTS + telephony surcharges.
- Document rollback — keep previous model id in config for instant revert.
Limits and open questions #
- No open weights — you cannot benchmark Tell-1 on-premises besidepplx-decider orSarvam Prism .
- Benchmark ownership — vCX is maintained by the vendor whose model now ranks first.
- Duplex stacks — Retell’s ownGPT-Live-1 changelog still serves callers who wantOpenAI full-duplex voice with a separate text planner; Tell-1 is a different architecture choice.
Related reading #
When to keep a frontier model on voice #
Keep Claude 5 Sonnet or GPT-6 Sol on Retell when your agent must cite long uploaded policies beyond Tell-1’s training distribution, when you need multilingual coverage Retell has not published for Tell-1, or when legal insists on a named foundation-model vendor in contracts. Tell-1 is a bet that phone-native finetuning beats general reasoning on median calls — not that frontier models lose on every row of the vCX board.
Pricing and benchmark claims reflect Retell’s public pages as of November 13, 2026. Confirm rates on retellai.com before budgeting.