Mira Insights | H1'26
How people actually use an AI agent when it lives in their messenger
Insights into six months of Mira – delegation habits, group-chat behavior, and the honest patterns behind hundreds of thousands of users.
REAL TASKS OUR USERS SOLVEBeyond chatting: The work people hand Mira
Almost everyone chats with Mira. The interesting part is what they do next.
- ~everyoneEveryday conversation & questions
- 29%Live web search for current answers
- 8%Creating & editing images
- 3%Generating video
- 2%Reminders & recurring automations
- 0.4%Connecting external services (Gmail, Calendar, Sheets)
Reminders remain the smallest slice of what people do with Mira. We think the bottleneck is onboarding, not demand. Tool connections tell the same story: most people lean on search and media, and only about 1 in 250 active users connects to an external service like Gmail or Calendar. Yet those connections are exactly what unlock Mira's most powerful, action-taking use cases.
WHAT GETS CREATED INSIDE TELEGRAM310,000 images, videos and tracks, and still counting
AI image generation leads by far, but the format mix keeps broadening.
- Images81%
- Video13%
- Music6%
All of it is made right in the chat, with zero setup. No separate creative app, no export step.
Creators don't stop at generating. They connect the tools they publish with: Canva and Figma for the visuals, Google Slides for decks, Instagram and YouTube for distribution, and Discord to keep the community together.
FROM AI ASSISTANT TO AI OPERATORThe real story of H1: Mira doesn't just answer, it acts
The most powerful use case is connecting Mira to external tools and letting it act inside them.
| Tool | What it's for |
|---|---|
| Gmail | Email triage & drafting |
| Google Calendar | Scheduling & events |
| Google Sheets | Data & tracking |
| GitHub | Code & repos |
| Google Drive | Files & storage |
WHAT PEOPLE TEACH MIRA TO DO40K+ people have already built at least one skill
A skill is a routine you describe once, and Mira runs it from then on.
What's interesting is that people aren't adopting templates. They're creating their own skills. Over the last 30 days, 99.9% of active skills were user-created rather than built by us.
And what they're creating isn't mostly work. It's routine. Motivation and morning briefs lead by a wide margin. Then come habit trackers, tarot readings, birthday reminders, and language lessons. Email triage is the only work-related skill in the top list.
GROUPS IN TELEGRAMMira isn't just a 1:1 assistant anymore
People have added Mira to 175,000 group chats, from family catch-ups to communities to business spaces. Friends and family make up the biggest share by far at 67.9%, followed by communities at 20.7% and business chats at 11.4%.
Inside those chats, Mira settles debates, answers the group's questions, cracks jokes, and breaks the ice with games. It acts less like a tool and more like a teammate everyone can call on.
RETENTION & ENGAGEMENT DEPTHThe clearest sign of a power user? They talk, not type.
NOT ONE FRONTIER MODEL IN THE TOP-5What Mira actually runs on
Across 460 billion tokens processed on OpenRouter, enough to make Mira a top-3 personal AI agent worldwide, the biggest names in AI barely show up. No GPT-5, no Claude, no Gemini anywhere near the top. Instead, roughly 95% of everything Mira does runs on open-weight models.
The media tells the same story. Most images come from ByteDance's Seedream and OpenAI's gpt-image-2, with Google's Nano Banana close behind. And for video, Mira leans heavily on the open-weight WAN 2.2, ahead of both Kling and Seedance.
WHAT'S NEXTThe trends we think will define H2 2026
Not something we measured. This is our bet on where agentic AI is headed, and where we're placing our own: agents that don't just talk, but act, pay, and coordinate.
💳 The agentic wallet
Agents are starting to hold and spend money: Google's Agent Payments Protocol, Visa and Mastercard agent checkout, stablecoin rails. On TON, Mira's next step is an in-chat wallet, so it won't just find the thing, it'll buy it.
🧩 From answers to actions
We're betting the next unlock isn't smarter replies. It's chaining more actions per request: Mira's value grows from depth of automation, not from any single answer getting cleverer.
🔗 Agents that work together
Standards like MCP are turning lone AI agents into networks. Our roadmap: agents on TON that share memory and pass tasks to each other.
🌙 Ambient & proactive
Reminders are still niche today, about 2% of tasks. We think that's an onboarding gap, not a lack of demand. Our bet: fix discovery with better onboarding and more educational content, and this becomes the default.
A note on the data
Every figure here is aggregated across our user base in H1 2026. We never disclose individual user data or revenue.
