OpenRouter as a Window into Real AI Model Adoption
AI model adoption describes how developers and enterprises move from testing large language or image generation systems to routing real, recurring production traffic through those models based on measurable performance, cost, and reliability rather than hype or marketing claims. OpenRouter has become a key place to watch this shift, because it functions as a neutral marketplace where teams can send tokens across many popular AI models without being locked into a single provider. Its usage data highlights which systems carry day-to-day workloads instead of headline demos. Combined, the top ten companies on the platform process roughly 19 trillion tokens, with the top four alone handling more than half of all activity. That concentration reveals where practical value is being found, and which vendors are turning benchmarks into sustained, production-grade traffic.
DeepSeek, Anthropic and Google Lead the Token Race
OpenRouter usage data shows a clear pecking order among language model providers. DeepSeek tops the list with 3.1 trillion tokens and 16.3% share of volume, pushed by its V4-Pro line and sharp cost advantage in benchmarks like the Artificial Analysis Intelligence Index. Anthropic follows with 2.94 trillion tokens, or 15.5%, reflecting the pull of Claude models for complex reasoning, coding, and agentic tasks. Google holds third place at 2.51 trillion tokens and 13.2%, backed by the Gemini family’s presence in widely used products. One quotable takeaway is that “combined, the top 10 companies on the platform processed roughly 19 trillion tokens, with the top four alone accounting for more than half of all activity.” In practice, this means AI model adoption is consolidating around a handful of providers that can deliver both capability and predictable economics at scale.
Google’s Sweep of Image Generation Models
In image generation models, Google has moved from challenger to default choice on OpenRouter. The company’s Nano Banana lineup dominates traffic: Nano Banana (Gemini 2.5 Flash Image) leads with 1.71 million requests and a 40.7% share, Nano Banana 2 (Gemini 3.1 Flash Image) follows with 1.21 million requests and 28.8%, and Nano Banana Pro (Gemini 3 Pro Image) adds 825,000 requests and 19.6%. Together, they account for roughly 89% of all image generation traffic on the platform. That grip comes from clear positioning: Nano Banana for high-volume, speed-sensitive tasks, Nano Banana 2 for price-conscious text-to-image quality, and Nano Banana Pro for high-fidelity commercial work with SynthID watermarking. OpenAI’s GPT-5.4 Image 2 stands fourth at 118,000 requests and 2.8%, indicating that, for now, Google’s image tools are the popular AI models developers route to when they need consistent, scalable results.
Why Speed and Performance Now Decide Model Choice
Speed is becoming a deciding factor in AI model adoption, especially when workloads scale. Data from the Artificial Analysis index ranks the fastest AI models by tokens per second, and the spread is wide enough to reshape developer choices. GPT-oss 120B on a high-compute tier leads with 306 tokens per second, ahead of its smaller sibling GPT-oss 20B at 239 tokens per second. Google’s Gemini 3.5 Flash reaches 212 tokens per second, while Alibaba’s Qwen3.7 Max sits essentially level at 211 tokens per second. xAI’s Grok 4.3 on a high tier delivers 190 tokens per second, with GPT-5.4 Mini on extra-high close behind. These throughputs matter because companies are now comparing speed-per-dollar as closely as they compare benchmark scores. When a fast model also undercuts rivals on cost in large benchmarks, it is more likely to move from test projects into production routing.
From Benchmarks to Production: How Preferences Are Shifting
The gap between theoretical capability and practical AI model adoption is narrowing, but not disappearing. Providers with strong benchmark scores still need to prove they can run agentic workloads cheaply and reliably. DeepSeek’s V4-Pro, for instance, scores 1554 on GDPval-AA and costs USD 1,071 (approx. RM4,900) to run the Artificial Analysis Intelligence Index benchmark, compared to USD 4,811 (approx. RM22,100) for Claude Opus 4.7. According to Artificial Analysis, this makes it more than four times cheaper while staying competitive on performance, which helps explain why production platforms like Lindy switched entirely to DeepSeek V4 after systematic testing. On the image side, Google’s accidental Nano Banana naming moment turned into a self-reinforcing feedback loop of usage and improvement. Together, these patterns show developers routing traffic toward models that solve concrete problems with predictable latency and cost, not only those leading headline benchmarks.
