What 2025’s AI API Market Data Reveals About the Future of Large Models

01 Q1 Token Surge and Long‑Tail Demand

The first quarter of 2025 saw an almost four‑fold increase in total token usage on OpenRouter, stabilizing at around 2 T tokens per week, with no further significant growth thereafter. This burst was driven by a strong long‑tail demand, as many niche models collectively accounted for 600‑700 B tokens each week.

Gemini‑2.0‑Flash leads usage, followed by Claude‑Sonnet‑4 and Gemini‑2.5‑Flash‑Preview‑0520.

DeepSeek V3 (free and paid) together could rival the second‑place model in usage.

DeepSeek‑V3 maintains a top‑10 position with high user retention.

Gemini‑2.0‑Flash’s low price (US$0.4 per M tokens), high capacity, and speed keep it in the top three.

Gemini‑2.5‑Flash is gaining momentum and may overtake its predecessor as prices drop.

Claude‑3.5‑Sonnet completed its lifecycle in March; Claude‑3.7‑Sonnet is nearing its end.

Claude‑Sonnet‑4 now holds the market position of earlier Claude models with stable usage.

OpenAI models fail to maintain a consistent top‑10 weekly usage.

GPT‑4o‑mini shows volatile usage, especially a spike in May likely due to OpenAI marketing.

02 Google Gemini Leads Market Share

Google holds a commanding 43.1% share of the large‑model API market, followed by DeepSeek (19.6%) and Anthropic (18.4%). The data reveal several trends:

Google is aggressively encroaching on Anthropic’s share.

DeepSeek’s market share has been stable and continues to grow since the release of V3.

OpenAI’s share is highly volatile, ranking fourth but far behind Anthropic.

Llama’s share has shrunk to about one‑fifth of its peak.

All other models together account for less than 10% of the market.

Gryphe, once notable for fine‑tuned Llama‑2 models, has disappeared from the rankings.

03 Model Usage by Domain

Usage patterns differ markedly across application domains:

Programming: Claude‑Sonnet‑4 dominates with 44.5% share, followed by Gemini‑2.5‑Pro.

Text Translation: Gemini‑2.0‑Flash holds an overwhelming lead due to its volume, low cost, and speed; seven of the top models are Google’s, suggesting many translation tools default to Gemini.

Role‑Playing: The market is highly fragmented; niche models collectively hold 26.6% share, with DeepSeek leading thanks to its higher hallucination propensity, and Gemini‑2.0‑Flash in third place.

Marketing: GPT‑4o is the clear leader with 32.5% share, reflecting OpenAI’s effective training for non‑programming tasks.

04 API Interface Trends: Code‑Writing Tools Dominate

The most frequently used OpenRouter interfaces are geared toward code generation:

Cline – top usage, focused on code writing.

RooCode – second place, also for coding.

liteLLM – third, a routing library for building various applications.

KiloCode – fourth, another coding tool.

SillyTavern – fifth, a local LLM interface similar to Ollama.

05 Overall Observations

Key takeaways from the data:

Google now controls nearly half of the large‑model API market, offering cost‑effective solutions like Gemini‑2.0‑Flash.

Anthropic focuses on programming, with Claude‑3.5, Claude‑3.7, and Claude‑4 providing a smooth transition between versions.

OpenAI’s market performance is weak, possibly due to access‑key restrictions and pricing issues.

DeepSeek enjoys strong user stickiness; DeepSeek‑V3 outperforms the older V1, likely because V1’s latency hampers token output.

Meta’s Llama series continues to decline.

Mistral AI holds about 3% market share, mainly among European users who favor fine‑tuned open‑source models.

X‑AI’s Grok series shows limited market positioning and needs significant development to become a SOTA contender.

Qwen series captures 1.6% of the market, indicating room for growth.