The Short Answer
Duolingo says the AI inference behind one Video Call fell from about $0.30 to well under $0.01, mainly by moving to open-weight models. The saving came from choosing the model per task. On ibl.ai you own all the code and the data and run every agent model-agnostic, so that choice stays yours.
The figures come from Duolingo's own earnings calls: the Q2 2026 call on August 5, 2026, and the Q3 2026 call on November 3, 2026.
What did Duolingo say about its Video Call AI cost?
Duolingo's CEO, Luis von Ahn, said on the November 3 call that Video Call "was costing about $0.30 per call, just generally on AI inference." He added: "We've managed to reduce that cost to well under $0.01 per call."
He named the cause in the same answer: "that's mainly by moving to open-source models or open-weight models." That is a drop of more than 30 times, a cut of more than 96%.
The number is not new this week. Von Ahn first gave it on the August 5 call: "It is now under $0.01 per video call." The November call repeated it with more detail.
What does the short version of the Duolingo story leave out?
The short version leaves out two details, and neither changes the conclusion. First, "per call" here means one Video Call conversation with Duolingo's AI character, not a single request to a model.
Second, Duolingo did not switch off frontier models. On November 3 von Ahn said "There are some small cases where a frontier model's still needed."
In August he said the company wants "to use open-weight models as much as possible," because "sometimes the frontier models are more advanced."
So the accurate version is a mix: open-weight models for the bulk of the work, frontier models where they still earn their price.
Why did the cost of Duolingo's Video Call matter to its business?
The cost decided who could get the feature. At $0.30 a call, von Ahn said in August, Duolingo "decided to put it behind our most expensive plan, which is Max."
Once the cost fell, Video Call moved down a tier. On November 3 he said: "Video Call for the longest time was just available in Max. Now that it's in Super." The move also sits alongside Duolingo's plan, stated on the same call, to sunset Max.
The CFO, Gillian Munson, said "The AI costs have not been as high as we thought, which is great," and named "less AI costs than we expected" as a driver of gross margin. She also expects margin to dip slightly as Video Call rolls out to all paid subscribers.
Are open-weight models good enough for production AI features?
For most of Duolingo's users, the company says yes. Von Ahn's words on November 3: "for the majority of our users, open-weight models are about as good as frontier models."
Duolingo is not the only one reporting this. Revolut says self-hosting open GLM models cut its AI coding costs to roughly a tenth, about $0.30 per million tokens; we went through what Revolut's numbers actually say.
The pattern across both: most production work, conversation practice, completion, extraction, routine drafting, runs well on models that cost a fraction of the frontier. For how open weights now set the price of closed models, see the open-weight price floor.
How can an organization switch AI models without rebuilding its product?
An organization can switch only if the model is a setting, not a dependency. Duolingo could move because it was free to pick the model for each task.
On ibl.ai, each agent's LLM is chosen per agent. On the organization in our recording, the catalogue listed 45 model providers and 316 chat models on one grid; it syncs itself every six hours, and a switch keeps the agent's prompts, data and connectors.
Costs then shows what each agent spends, so the saving from a switch is measured rather than assumed.
Who should own the AI stack that makes a model switch possible?
The organization paying for the inference should own the AI stack. A product rented per seat from one model vendor cannot follow prices down; Duolingo's could.
On ibl.ai you own all the code and the data, the platform runs model-agnostic, there is no per-seat pricing, and you can deploy anywhere, from your own cloud to a fully air-gapped network.
That is the same freedom to choose that cut Duolingo's cost per call, without building the platform yourself.
ibl.ai is family-owned and operated from New York, NY — a U.S.-headquartered, domestically-owned long-term partner, not a vendor that sells licenses and moves on.
Sources: Duolingo's Q3 2026 earnings call (November 3, 2026) and Q2 2026 earnings call (August 5, 2026), transcripts published by The Motley Fool. The brief called it "30x cheaper per call" and a switch to open-weight models; the calls support more than 30 times cheaper per Video Call conversation, first disclosed in August, with frontier models kept for some cases.