The disappearing model picker
Anuma's whole premise is choice: dozens of models from every major provider, one selector. So the most surprising trend in our own data is that people are choosing not to choose. In March, when the auto router launched, 11% of messages left the selector untouched. By May it was a majority. In July it is 87% of all messages.
Growth doesn't explain it away. The opposite. Monthly senders grew more than five-fold over the same stretch, and the number of people who manually picked a model fell in absolute terms, from about 2,200 in March to about 900 in July. The new users aren't learning the selector; the old users are letting go of it.
The obvious objection: isn't this just new users never touching a default? The data says no. Split each month's senders into new (first message that month) and returning, and both lines climb. New users' auto share rose from 13% in March to 92% in July, and returning users converted from 9% to roughly 60% before plateauing. That plateau is the tell: the veterans who remain manual are a committed core, not laggards. And the monthly gains are already shrinking (+27, +19, +18, then +11 points), the flattening shoulder of a classic adoption S-curve. No single release produced a jump anywhere in the series; this is trust accumulating, not a forced migration.
The 9% who still choose
In July, roughly 9% of active users ever touched the model selector. But they are not a residue: they are the heaviest users on the platform. The average chooser sent 16.9 messages in July against 4.3 for auto-only users (medians 8 and 3), and that 9% of people produced 28% of all message volume.
This reframes what a model leaderboard measures. Our Index Board counts deliberate picks precisely because they are deliberate: the considered judgment of the platform's power users, not the passive output of a default. Choice has become an enthusiast behavior, and enthusiast behavior is the leading indicator: it's where shifts like the Qwen surge show up weeks before they reach anyone else's charts.
How Anuma's auto decides
When a message arrives with the selector untouched, Anuma's backend classifies it, using lightweight rules over the message itself (never a stored profile), into one of five task buckets: vision (the message carries an image), image generation, coding, reasoning, or general conversation. Each bucket maps to a designated model, chosen for that task's strengths and updated by our team as the model market moves.
There are two separate mappings, and the split is the interesting part. In standard chats, as of late July, general messages route to GPT-5.6 Luna, coding and reasoning to Claude Sonnet 5, vision to Grok-4.5. In private mode, the pool is entirely open-weight: MiniMax M2.7 by default, Kimi K2.7 Code for coding, GLM-5.2 for reasoning. That's consistent with what our board shows about open-weight models carrying roughly a quarter of deliberate picks: they are no longer the budget tier, they are the privacy tier's first string.
What auto does not do matters too: routing considers the task, not the user's wallet (there is no cost or load input in the model choice), and the full model catalog stays one tap away. Auto is a default, not a wall.
The market fought the router — then adopted it
When OpenAI launched GPT-5 in August 2025, it removed the model picker entirely and forced every ChatGPT user through a router. The result was one of the most vocal user revolts in the product's history, and within days OpenAI restored the picker. Less than a year later the picture has inverted: Auto is ChatGPT's default and most users never change it. Underneath, routing became infrastructure — OpenRouter's Auto runs on NotDiamond, and LMSYS's RouteLLM turned the problem into a research benchmark.
Our data suggests the 2025 backlash was never really about routing. It was about removed choice. Anuma has kept the selector one tap away the entire time, and with choice fully available, adoption of auto reached 87% on its own. People don't object to a machine picking their model. They object to being told they can't.
Methodology
Message counts and auto share are from Anuma's client-side send events (metadata only, no prompt content) over calendar months February–July 2026, with July partial through the 27th. “Auto” means the message was sent with the model selector untouched. February predates the router's launch. The chooser/auto-only engagement split groups July's senders by whether they sent at least one manually-picked message that month. The cohort split classes a sender as “new” in the month of their first-ever message and “returning” afterward. Frontend events are ad-blocker-sampled, and our user base is wallet-gated and mobile-heavy: directional, not a general-market sample. The routing description reflects Anuma's production configuration as of late July 2026; the task-to-model mapping changes as models ship.