CodeSOTAIntelligence
Demand Research · ort.fabryka.ai

Who Burns the Tokens: OpenRouter Apps

Models are the supply side; apps are the demand. This maps the ~45 ranked apps to the models and vendors they run on — revealing how concentrated each vendor's volume is on a handful of agent apps.

snapshot 2026-08-05 · top-20 model breakdown per app · generated 2026-08-05 · ← dashboard · market memo

Token flow — apps → vendors → origin

Where the demand actually goes: each top app's tokens flowing to the vendors it runs on, then to the model's country of origin. Ribbon width = tokens; red = Chinese-origin, blue = Western.

Hermes AgentKilo CodeClaude CodeOpenClawClinepiDescriptISEKAI ZEROJanitor AILemonadedeepseektencentanthropicstepfunOther vendorsz-ainvidiaminimaxpoolsideChinaWestOther

Token flow — apps → the actual models

The same demand, resolved to the concrete models each app routes to (top 9 models; the rest grouped as "Other models").

Hermes AgentKilo CodeClaude CodeOpenClawClinepiDescriptISEKAI ZEROJanitor AILemonadeOther modelsHy3DeepSeek V4 Flash 0423Step 3.7 FlashGLM 5.2DeepSeek V4 ProNemotron 3 Ultra (free)MiniMax M3Claude Sonnet 5DeepSeek V4 Flash 0731

The numbers behind the flow

Top apps by token volume (T)
Demand mix per app — 🔴 Chinese · 🔵 Western · ⚪ other (% of the app's tokens)
Vendor concentration — top-app share of each vendor's app volume (red = ≥40% concentration risk)

1 · App leaderboard — who consumes the most

apptokensshareruns on (vendor mix)
Hermes Agent28.3T46.2%deepseek 31% · tencent 21% · nvidia 11%
Kilo Code7.1T11.6%stepfun 27% · tencent 19% · poolside 14%
Claude Code6.8T11.1%z-ai 41% · anthropic 36% · deepseek 7%
OpenClaw3.5T5.7%minimax 25% · deepseek 22% · anthropic 12%
Cline3.0T4.8%stepfun 26% · poolside 24% · deepseek 23%
pi2.1T3.5%deepseek 34% · anthropic 18% · z-ai 15%
Descript1.9T3.0%anthropic 87% · google 11% · openai 1%
ISEKAI ZERO945.9B1.5%deepseek 59% · xiaomi 30% · z-ai 4%
Janitor AI817.9B1.3%deepseek 65% · nvidia 13% · z-ai 8%
Lemonade805.6B1.3%google 36% · tencent 32% · openai 17%
Codex767.3B1.3%openai 85% · deepseek 6% · tencent 2%
Pieces.app: Ambient Artificial496.6B0.8%google 75% · x-ai 24% · anthropic 1%
Portkey AI469.8B0.8%google 36% · deepseek 32% · z-ai 9%
Nous Research API467.6B0.8%stepfun 31% · deepseek 28% · tencent 21%
Stanley413.4B0.7%minimax 51% · anthropic 45% · google 4%
Mira is the leading AI agent i411.4B0.7%minimax 51% · openai 28% · z-ai 14%
LangChain353.4B0.6%google 39% · deepseek 26% · anthropic 17%
GDevelop349.3B0.6%openai 44% · deepseek 43% · google 12%
Zed Editor306.3B0.5%deepseek 27% · z-ai 19% · xiaomi 13%
SillyTavern227.9B0.4%deepseek 42% · z-ai 22% · anthropic 18%
extra.email221.2B0.4%deepseek 98% · qwen 2% · anthropic 0%
HammerAI191.7B0.3%mistralai 93% · sao10k 5% · aion-labs 1%
Craft189.7B0.3%google 39% · openai 24% · xiaomi 20%
Ito184.0B0.3%openai 99% · anthropic 1% · moonshotai 0%
OpenHands161.2B0.3%z-ai 30% · deepseek 17% · anthropic 16%

Share = % of tokens across all tracked apps. "Runs on" = the app's top vendors by token volume.

2 · Vendor demand concentration — who depends on whom

vendortokens (via apps)top-app conc.top driving apps
deepseek14.2T62%Hermes Agent 62% · OpenClaw 5% · pi 5%
tencent8.6T70%Hermes Agent 70% · Kilo Code 15% · Claude Code 4%
anthropic6.7T36%Claude Code 36% · Descript 24% · Hermes Agent 13%
z-ai5.9T47%Claude Code 47% · Hermes Agent 31% · pi 5%
stepfun5.9T45%Hermes Agent 45% · Kilo Code 33% · Cline 13%
nvidia4.7T64%Hermes Agent 64% · Kilo Code 15% · OpenClaw 7%
minimax4.1T49%Hermes Agent 49% · OpenClaw 22% · Kilo Code 14%
poolside2.6T39%Kilo Code 39% · Hermes Agent 30% · Cline 27%
openai2.0T32%Codex 32% · Hermes Agent 13% · pi 10%
xiaomi2.0T34%Hermes Agent 34% · Kilo Code 17% · ISEKAI ZERO 14%

Top-app conc. = share of the vendor's app-driven volume from its single biggest app — high (≥40%) = concentration risk, low (<20%) = broad, durable demand.

3 · Why it matters