Observations on Large Model Usage in the First Half of 2025: Gemini Series Accounts for Half of the Market Share, DeepSeek V3 Users Retain Very Highly

Halfway through 2025, text generationLarge ModelIs it already in the second half? Is OpenAI completely ignoring the API market? Is Grok3 not used at all? What is the future of the "Great Model War"?

Recently, Twitter blogger "karminski-dentist" released a post about the first half of 2025 big model API market data analysis. Based on OpenRouter data, "karminski-dentist" analyzed the first half of the year big model's total Token usage ranking and trends, the market share share of different big models, the application preference of segmented models, and the trend of API interface usage, and came up with some very interesting observations. Below are the details of the analysis.

Source:https://x.com/karminski3/status/1942612077241311386

01,Total AI Token usage grew nearly fourfold in the first quarter, andremainEnlivening long-tail demand

First, let's take a look at the trends of the most popular models. The latest data shows that Gemini-2.0-Flash is in first place, followed by Claude-Sonnet-4 and then Gemini-2.5-Flash-Preview-0520.

Coming in at #4 and #5 are the free and paid versions of DeepSeek V3 0324, respectively, and if you add the usage of both together, DeepSeek-V3 may reach the second place level of usage.

In addition, there are some unique observations that we can draw from the trend graphs:

• The first quarter of 2025 saw an absolute explosion of AI growth, with OpenRouter's total Token usage quadrupling in the first quarter of 2025 compared to the previous quarter, and then stabilizing at 2T Token per week. There has been no significant growth since then.
• The usage of other models stabilized at 600-700B Token usage after exploding in the first quarter. This situation reflects the diversity of market demand to a certain extent, and the huge long-tail volume proves that the model market is active and there is segmented demand.
• DeepSeek-V3 has been stable in the Top 10 since its release and has an extremely high user retention rate.
• Gemini-2.0-Flash has remained in the top three models in terms of usage because of its low pricing ($0.4 per million tokens exported), large capacity, and speed.
• Gemini-2.5-Flash is gaining momentum, and Google's modeling strategy is well-positioned considering that it will likely replace Gemini-2.0-Flash when the price drops.
• Gemini-2.5-Pro replaced the previous experimental version, but there was no significant increase in usage.
• The Claude-3.5-Sonnet completed its historic mission at the end of March this year, and the Claude-3.7-Sonnet is nearing the end of its life cycle.
• The Claude-Sonnet-4 has now taken over the market place of the previous Claude line of models, but its use has remained steady without sustained significant growth.
• OpenAI's models do not guarantee that one of them will remain in the Top 10 consistently in terms of weekly usage.
• GPT-4o-mini usage has fluctuated widely, with a particularly strong showing in May, likely stemming from the results of OpenAI marketing.

02,Google's Gemini series is firmly established as the No. 1 in market share, theOpenAI Fluctuates Significantly

In terms of market share, Google holds the first place with 43.11 TP4T share, DeepSeek and Anthropic are in the second and third place with 19.61 TP4T and 18.41 TP4T share share respectively.

From the market share data, we find:

• Google is now strongly squeezing the market share belonging to Anthropic.
• DeepSeek has maintained a market share since the release of DeepSeek-V3 and continues to grow.
• OpenAI's share fluctuates particularly sharply, with a significant gap to Anthropic in the top spot, despite being ranked fourth.
• Llama's share continues to shrink, having shrunk to about one-fifth of its peak.
• The total share of other models does not exceed 10%.
• Gryphe, an organization focused on fine-tuning models, has disappeared from the rankings, and Gryphe's MythoMax13B model, which was fine-tuned based on the llama2 model, was at one time particularly popular in AI role-playing scenarios.

03,Four giants in different segments

In terms of usage data for the segmentation model, we find:

• In the field of programming, theClaude-Sonnet-4 is the clear leader with 44.51 TP4T share, followed by Gemini-2.5-Pro.
• In the field of text translation.Gemini-2.0-Flash's overwhelming dominance stems largely from its high usage, affordability and speed. Another surprising finding is that seven of the top models in the ranking are Google's models, except for the second model that occupies 20% share. It is speculated that some translation software may have integrated Google models by default.
• In the area of role-playing.The market shows a high degree of fragmentation, with niche models combining for a 26.61 TP4T share. Next is DeepSeek leading the way in roleplaying with its high hallucinatory tendencies. Third place goes to Gemini-2.0-Flash with its affordable price and high usage.
• In the field of marketing.GPT-4o is the undisputed and absolute leader with a 32.51 TP4T share, which probably reflects the fact that OpenAI is quite effective at training in non-programming specialties and that users really like GPT-4o's output.

The amount of model calls in the programming domain

The amount of model calls in the translation domain

The amount of model calls in the role-playing domain

The amount of model calls in the marketing domain

04,API Trends in interface usage:Coding tools dominate

Finally, take a look at what interfaces (interfaces) people mainly use on OpenRouter:

• Rounding out the top two are Cline and RooCode, both of which are primarily used for writing code.
• Third place goes to liteLLM, a routing library for building a variety of applications.
• Fourth place KiloCode, also for writing code.
• Fifth place went to SillyTavern, an Ollama-like native Large Language Model (LLM) interface through which to connect and interact with the Big Model.

05,General Observations

Based on these data observations, we draw several conclusions:

• Currently, Google has almost half of the big model API market, and its solution models cover a wide range of domains, even including the affordable, high-value Gemini-2.0-Flash (which is cheaper than DeepSeek).
• Anthropic, on the other hand, focuses on programming, and its Claude-3.5, Claude-3.7, and Claude-4 models provide a smooth transition between old and new versions.
• OpenAI's performance in the large model API market has not been strong, possibly due to a variety of constraints, such as the need to apply for an AccessKey on their official website for the latest version of the model, or pricing issues.
• DeepSeek models have strong user stickiness. Surprisingly, DeepSeek-V3 is the most popular in the market rather than the DeepSeek-R1, probably because DeepSeek-R1 may take too long to process and the first valid Token output is too slow, resulting in less users than V3.
• Meta's Llama family of models is in decline.
• Mistral AI's models have a surprisingly high market share of about 3%, and my personal exposure to Mistral AI users has been relatively limited, mainly to European users who like to fine-tune open-source models.
• X-AI's Grok line of models has made some progress, but the market niche is unclear. If X-AI's goal is to become a SOTA model, they have a long way to go.
• The Tongyi Qianwen (Qwen) series of models captured a market share of 1.61 TP4T and needs to continue its efforts.