Key Intelligence Insight
Mirage (formerly Captions) builds foundation models for talking video and operates two commercial products: the Captions consumer app (20 million users) and Mirage Studio, a text–to–video enterprise platform for brands. Founded in 2021 by Gaurav Misra, a former Snap design engineering lead; rebranded September 2025; $100 million raised at a $500 million valuation from Kleiner Perkins, Sequoia, and Andreessen Horowitz. The company's technical differentiation is architectural: competitors in the avatar space (Synthesia, D–ID, Hour One) use neural rendering pipelines that require per–subject training and plateau at a ceiling of semi–realism.
Mirage is building large diffusion models for talking video from scratch, with no pre–existing footage, no voice cloning, and no lip–sync overlay. Misra's framing for why this is winnable: media generation is a bounded problem, with a defined ceiling of photorealism, unlike the open–ended problem of intelligence that LLMs attack. That ceiling can be reached and solved.
The product strategy compounds this technical approach with a consumer flywheel: Captions app revenue funds the foundation model research; the research produces generation–four capabilities (unlimited–length video, multi–person scenes, simultaneous audio–video generation); those capabilities feed back into both the consumer app and the enterprise studio. The go–to–market is explicitly non–competitive with professional editing tools – Misra positions Captions against Canva, not Premiere Pro, targeting the business user who needs video done for them, not the professional who wants control.
Founding Story
Gaurav Misra, CEO and co–founder, built the design engineering function at Snap: a prototyping team embedded inside a 3,000–person engineering organization, tasked with testing product concepts at limited real–world scale before committing production resources. That experience produced two durable operating instincts – ship a marketable feature every week, and maintain a secret roadmap alongside the public one.
Captions launched in 2021. The product grew to 500,000 users and $500,000 in monthly recurring revenue over 18 months during which the founding team was focused elsewhere and released no updates. That growth through neglect confirmed product–market fit at a level Misra describes as rare. It also revealed the ceiling: the creator market was enormous, the skill gap it was eliminating was real, and the underlying model research required to push realism to its limit was not yet underway. In September 2025 the company rebranded as Mirage, a name chosen to reflect the ambition: full–stack foundation models for video, not a feature layer sitting on top of someone else's infrastructure.
Product
Mirage operates two commercial surfaces.
The Captions consumer app serves over 20 million content creators with AI–powered recording, editing, captioning, and publishing tools optimized for TikTok, Reels, and YouTube Shorts. The eye contact correction feature – which redirects a speaker's gaze toward the camera during recording, built in partnership with NVIDIA – is the canonical product expression of Mirage's secret roadmap approach. No user requested it. It went viral across dozens of languages, shifted creator behavior permanently, and has since been replicated across the industry.
Mirage Studio, launched June 2025, targets brands and advertisers. A user provides a script or audio file, generates a photorealistic AI avatar with body language and facial expressions calibrated to the script's emotional register, specifies a background via text or image upload, and receives a finished talking–head video. Nothing requires a camera, a studio, or talent. Multiple takes of each line can be generated and selected, giving the user directorial control over gesture and delivery. The brand owns the avatar as IP.
Market, Competition & Business Performance
Market
The talking video problem is the largest unsolved media generation problem remaining. Talking–head video – a person, speaking to camera, with natural expression, gesture, and audio – constitutes the dominant format of social media, advertising, e–learning, customer communications, and political messaging. It has historically required a camera, a subject, a production environment, and post–production resources.
The available market for removing those requirements is not niche. It is every business that currently cannot afford to participate in video, every brand that needs 40 ad variations instead of one, every company that needs to localize a video into 20 languages without booking a studio for each. Misra frames the historical trajectory clearly: websites were once optional, then expected, then essential. Video is following the same curve, and the businesses that cannot produce it are being progressively excluded from the primary discovery channels for products and services.
Competition
Two categories of AI video companies currently exist, and both have structural ceilings.
B–roll generators – Runway, Pika, Luma Labs – produce visually compelling silent footage. They do not generate speech, expression, or the human–to–camera interaction that constitutes most of what is actually watched on social media. The mechanism is scenographic, not communicative.
Avatar platforms – Synthesia, D–ID, Hour One – generate talking–head video by training on real subjects, applying face–swap or lip–sync technology on top of pre–recorded body doubles. The ceiling is architectural: neural rendering pipelines cannot generalize to arbitrary subjects, cannot produce the expressive range that makes a speaker believable, and cannot scale to the realism levels that large diffusion models are approaching.
Mirage is building in neither category. Its foundation model treats talking video as a generative problem from scratch – no pre–existing footage, no voice cloning, no lip–sync overlay. The model generates a photorealistic person, with body language calibrated to the emotional register of each line, speaking in continuous expressive video of arbitrary length. As that capability matures, avatar tools built on older infrastructure face a realism gap they cannot close without architectural replacement. That is the displacement mechanism.
Business Model
Mirage runs a dual–tier subscription model. The Captions consumer app starts at $25 per month. Mirage Studio targets brands at $399 per month for 8,000 credits. The enterprise positioning is brand IP creation: businesses own the likeness of their AI representative outright, eliminating the variable cost structure of human talent production – no usage rights, no per–project negotiations, no contract overhead.
Traction
Over 20 million users on the consumer platform. More than $100 million raised, including a $60 million Series C at a $500 million valuation, backed by Kleiner Perkins, Sequoia Capital, and Andreessen Horowitz. Forbes AI 50 and Fast Company Most Innovative Companies in 2025. The company employs 51 to 200 people, with 136 LinkedIn–associated members and 6% annual headcount growth. Operations have grown 50% year–over–year. The Mirage app is at mirage.app.
The foundation model bet is credible. The commercial surface is performing. The open variable is whether Mirage can reach fourth–generation quality – unlimited–length video, multi–person scenes, simultaneous audio–video generation – before a better–resourced competitor closes the same gap, and whether the safety infrastructure required to deploy that capability responsibly can be built as fast as the capability itself.