CodeSOTAIntelligence
CodeSOTA · Data Memo

The Great Model Migration

Every month, billions of tokens quietly change hands between AI labs. The flows reveal who is winning developer demand — and who is being substituted away.

ort.fabryka.ai · August 05, 2026 · 150-day window · see normalized to market share →

There is a comfortable story that the AI market has settled into a stable oligopoly. The usage data on OpenRouter — the largest neutral router of LLM traffic — tells a different one. Share is moving, fast, and in a consistent direction.

Vendor share of OpenRouter tokens (%, 7-day average)

The lines cross. That is the whole point. A market in equilibrium produces flat, parallel bands; this one produces a reshuffling. Over the window, Deepseek added +19.6pp of share while Minimax shed -12.7pp.

The substitution is directional

Share shifts are not random churn. When you rank vendors by net change, the market splits cleanly into two camps — and the gainers' rise tracks the losers' decline almost day for day.

Net share shift over the window (percentage points)

Who is taking whose lunch

Pairing vendors by anti-correlated share trajectories surfaces the cleanest substitutions — one rises precisely as the other fades:

The numbers are abstract; the pictures are not. Each panel overlays one rising vendor on one fading vendor — the same share, changing hands:

Deepseek rising as X-Ai falls · corr -0.83
Deepseek rising as Google falls · corr -0.74
Tencent rising as X-Ai falls · corr -0.73
Xiaomi rising as Qwen falls · corr -0.63
Xiaomi rising as Google falls · corr -0.57
Xiaomi rising as Anthropic falls · corr -0.53

Gaining: Deepseek, Xiaomi, Tencent. Fading: Minimax, Google, Anthropic.

The pattern is not subtle: cheaper, open-weight challengers are absorbing the volume that incumbents like Google and Anthropic once took for granted. Volume is the leading indicator; revenue is the one that moves last.

Inside each lab, an upgrade treadmill

Zoom into any single vendor and a second migration appears — between its own models. Usage doesn't sit still on a version; it climbs the release ladder. Each band below is one model; watch how the newest release swells as the prior one thins.

Anthropic runs mostly on Claude Sonnet 4.6 today (19% of its volume).
Google runs mostly on Gemini 2.5 Flash Lite today (15% of its volume).
Deepseek runs mostly on DeepSeek V4 Flash 0731 today (43% of its volume).
Openai runs mostly on GPT-5.6 Luna today (56% of its volume).
Xiaomi runs mostly on MiMo-V2.5 today (93% of its volume).
Minimax runs mostly on MiniMax M3 today (89% of its volume).
Z-Ai runs mostly on GLM 5.2 today (91% of its volume).
X-Ai runs mostly on Grok 4.5 today (73% of its volume).
Tencent runs mostly on Hy3 today (99% of its volume).

This is the incumbent's consolation prize: in-family migration is retention, not churn. A user moving from Opus 4.6 to 4.7 still pays the same lab. The danger is the other chart — when the next upgrade a developer reaches for carries a different logo.

What to watch