Retell's $100M Run Rate Came With a Model: What Tell 1 Actually Is

On 12 November 2026 Retell published two things that will be read together whether or not the company intended it: a blog post announcing that it had crossed $100 million in annualized run rate, and, in the same week, the first public evidence of Tell 1 — a model Retell built itself and describes as its own, listed on the company homepage under "Developed by Retell Labs" as the reasoning layer of its voice stack. The revenue number is the one that travels. The model is the one that changes what Retell is, because a voice-agent platform that merely routes your prompt to G​PT or Claude is a different business from one that ships the brain as well.

What Tell 1 is not, as of this writing, is documented. There is no launch post, no changelog entry, no model card, no API, no pricing line and no weight release anywhere on Retell's site that names it beyond a label. The $100M post itself never mentions it. The only Retell-published page that says anything substantive is a benchmark changelog from 29 September 2026, carrying a November update: Tell 1 "ranks #1 overall on the Retell vCX Benchmark," leads on colloquial speech and latency, and loses to other models on grounding, instruction following and tool calling. That is a company ranking its own model first on a benchmark it wrote, and it says so in a sentence — which is more candour than most vendors manage, and still not evidence you can build a procurement decision on.

The two numbers, and what each one is worth

The revenue figure is Retell's own and has the shape of a private-company milestone rather than an audited result: $100 million annualized, reached on $4.6 million raised, in under three years, across 4,000-plus customers running more than 100 million calls a month. CEO Bing Wu put the trajectory at $10 million to $100 million in 17 months and added that the company is profitable. None of that is independently verified, and "annualized run rate" is a forward-looking construction — it multiplies a recent month, it does not report a year. Treat it as a strong signal about adoption, which the customer references do support, and not as a financial statement.

The model claim is harder. The widely repeated detail — that Tell 1 was trained on more than 1 million hours of customer calls — is attributed to Retell in secondary coverage, and I could not find it on any Retell-owned page: not the blog, not the changelog, not the docs index, not the model benchmark site. That does not make it false. It does mean that if the training-corpus figure matters to your evaluation — and for a contact-center buyer worried about where the model's behaviour came from, it should — you are currently taking a second-hand quote of a company statement. A million hours is roughly 114 years of audio. It is a plausible number for a platform running 100 million calls a month, and plausibility is not documentation.

What is documented, in machine-readable form, is the benchmark. Retell's vCX site ships its results as structured data, and the Tell 1 row reads: 88.6% instruction following, 97.2% on the row Retell labels hallucination, 95.4% tool calling, 96.6% on the colloquial-dimension score, for an average of 94.45% across 62 evaluated configurations from 45 distinct models. Median latency is 72.5 ms and cost is 3.5 cents per minute, which puts it first of 62 on latency and 19th on cost.

The caveat inside the leaderboard

Read the row before the rank. The Tell 1 entry was scored on benchmark version 17. Fifty of the 62 rows in that same table were scored on version 1. Only eleven rows sit on any version above 1 — the newest GPT-6.1 and G​emini-4 entries, a couple of Claude builds, and Tell 1. A benchmark that evolves its test set is doing the right thing; the problem is that a "#1 overall" computed across rows from different test versions is not a like-for-like ranking, and the page does not adjust for it.

Two of the four dimensions are also closer than the headline suggests. Tell 1's tool-calling score of 95.4% ranks fifth, and its instruction-following 88.6% ranks second behind GPT-6-Astra's 89.7%. Where it is genuinely out in front is the dimension Retell built the benchmark around: colloquial speech, at 96.6%, first place, and latency, where 72 ms is not a rounding difference from the 550-1,400 ms band most of the rows occupy. If you believe the harness, the claim that survives scrutiny is narrow and specific — this model is fast and it sounds like a person — and it is exactly the pair of properties a phone agent is judged on by a caller who does not know a model exists.

Why a platform builds its own model at all

Retell's own architecture explains the incentive. The homepage splits the stack into three steps and puts Tell 1 in the middle one, "Reasoning & orchestration," next to an agent harness that brings together models, knowledge and tools. In that position a model is not a product you buy; it is a component whose cost, latency and conversational register you control. A platform paying a frontier lab per token on every one of 100 million monthly calls has a permanent margin problem and a permanent latency floor, because the round trip to a third-party inference stack is inside the caller's pause. Owning the middle layer removes both, at the cost of having to be good enough at the parts you took over.

That is also why the gaps Retell admits to are the ones it can afford. Grounding, instruction following and tool calling are where a general frontier model still wins, and they are the dimensions a customer's own prompt, knowledge base and function definitions do most of the work on. The dimensions Retell kept for itself are the ones no prompt can fix: how fast the first syllable arrives, and whether the sentence sounds like a call centre script or a colleague.

The pricing that reaches the buyer is still per minute, not per token: Retell lists voice agents at $0.07 to $0.31 per minute as of November 2026, a range that bundles telephony, speech synthesis and the model. Tell 1 does not appear as a line item you can select or decline, which is the practical difference between a model and a platform feature. You cannot bring your own Tell 1, and you cannot take it anywhere else.

What this means if you are buying, not building

For an enterprise contact centre, the honest summary is that Tell 1 is a reason to re-test Retell and not yet a reason to standardise on it. Three things would change that, and none has happened. A model card with the training-data provenance, the context window and the tool-calling contract written down. A published price for the model separately from the per-minute bundle, so the cost can be compared against a bring-your-own-LLM configuration. And a second-party evaluation — Artificial Analysis, a public voice benchmark, or your own call recordings run through both configurations — because a vendor ranking itself first on its own harness is a starting hypothesis, not a result.

If you are testing, the test is cheap and specific: take the calls your agents actually handle, and compare Tell 1 inside Retell against the same Retell agent pointed at a frontier model you already trust. The benchmark predicts one outcome — Tell 1 will feel faster and more natural, and will make more grounding mistakes. Whether that trade is good depends entirely on whether your calls are lookups with a right answer or conversations with a nervous person, and no leaderboard can tell you which one you have.

What to watch

The most informative next development is not a better score. It is any of the things Retell has so far not published: the model card, the training-hours claim in its own words, a per-model price, or an independent run. A company at a $100 million run rate with 4,000 customers has every reason to publish all four, and the fact that a benchmark update note is currently the entire public record suggests the model shipped into production before it shipped into documentation. That is normal for infrastructure and unusual for a model, and it is the thing to re-check before you cite any Tell 1 number, including the ones above, as settled.