Consumer AI Model Rankings
Ranked from how people really use AI through Anuma — the models they pick, switch to, and stick with across every major provider, plus independent research on the shifts.
Where users defect
Of 1,258 users who tried more than one model, we tracked 4,471 switches between consecutive messages. The signature Anuma metric: not which model scores highest on a benchmark, but which one a real person leaves — and where they go next.
How people use Anuma
Two honest layers, both read from metadata — never a word of the prompt. First, the feature people pick in the composer. Then, the tools the assistant reaches for to answer — web search, weather, and the rest — which fire silently inside a normal chat.
On Modes, Anuma is still ~80% plain chat; the tell is in the tail — people reach for Opus 4.7 in Deep research and GPT-5.2 in Council. Flip to Tools and the picture the modes can't show appears: web search is the single most-invoked tool — ~48% of all tool calls — even though it fires silently inside an ordinary chat.
Different denominators: Modes is a share of all messages; Tools is a share of the ~6% of messages that fire a tool — so the two aren't meant to be read against each other.
Analysis & deep-dives
The research behind the numbers — data insights on what's moving, technical write-ups on how Anuma works, and perspective pieces on where the model market is going.
Inside Anuma's PII redaction
As Europe argues over scanning private messages, how Anuma strips personal data from prompts on-device — and what our own usage data shows about privacy in practice.
Read the analysis →The Qwen Surge — and sudden sunset
Qwen-3.6-Max rocketed from 4% of deliberate picks to the #2 spot in a month — then fell to near-zero the next week when it was retired. A case study in how fast consumer model preference moves.
Read →Where users defect: the models people quietly leave
1,258 users, 4,471 switches — the models losing the most users, and who picks them up.
Coming soonThe case for model-agnostic AI
Why betting on one model is the wrong call — and what routing across all of them unlocks.
Coming soonCouncil Mode, explained
Ask several models the same question and compare — how it works, and how people actually use it.
Coming soonHow the index is built
Two data streams — one for what people choose, one for what the platform serves — plus five rules that bound what the numbers mean.
- 01Deliberate choices only
- The auto router — now most messages — is resolved server-side, so it is excluded. This measures choice, not routing.
- 02Human, not agents
- 100% human sends from the app. Agents, background jobs and embeddings are excluded from rankings — but still counted in platform token totals.
- 03Usage, not quality
- We report what people chose, not benchmark scores. Plan and credit gating shapes the mix.
- 04Names canonicalized
- Hundreds of raw model strings collapse to real models before counting. Δ 90d compares current share to the 90-day baseline.
- 05Population
- A wallet-gated, mobile-heavy base — directional and early, not a general-market sample.