The AI landscape is evolving at an unprecedented pace, yet understanding demand and usage patterns, beyond standardized benchmarks or leaderboard platforms, remains a challenge. Meanwhile, the preferred model one week can easily shift with the introduction of a powerful upgrade from a frontier provider or an unexpected disruptor.
Our goal is to make Poe the best place to explore, compare, and harness the outputs of AI models. Since Poe users have provider-agnostic access to the latest frontier models in a single interface, underlying trends among them may herald broader shifts in the AI ecosystem.
Building off our previous report, this analysis displays weekly aggregated usage data from January 2025 to May 2025 among Poe users in several key, but expanded, domains: text, reasoning, image, video and audio. This includes the sustained growth of reasoning models following DeepSeek's viral moment, how image and video generation are becoming increasingly competitive, and early signs of diversification in audio.
We hope our latest findings offer researchers and the public a helpful glimpse into the rapidly expanding AI ecosystem. [1] [2]
Frontier labs are rapidly releasing smarter general-purpose text models [3]
OpenAI’s GPT-4.1 family and Google’s Gemini 2.5 Pro, both of which offer improved performance on real-world coding tasks, rapidly increased to message shares of ~10% and ~5%, respectively, within weeks of launch.
Anthropic’s Claude family (e.g. Claude 3.5 Sonnet and Claude 3.7 Sonnet) saw a ~10% absolute decline in share over the same period.
DeepSeek's viral moment appears to have waned, as other affordable, verbose reasoning models have been released, with DeepSeek R1’s message share declining from a peak of 7% in mid-February to 3% by the end of April.
Similar to the findings in the previous report, new flagship models within an individual provider’s offering appear to cannibalize their predecessors. In this case, Poe subscribers rapidly adopted Claude-3.7-Sonnet over Claude-3.5-Sonnet, although the latter retained a notable ~12% overall usage among LLMs.
Reasoning models sustain usage following DeepSeek’s viral entry earlier this year
Since the start of 2025, frontier labs have been rapidly iterating on their reasoning model offerings. This has resulted in an increase in models that are capable of spending more time and compute to solve complex tasks with more precision and reliability. Notably, the share of all text messages sent to reasoning models on Poe grew from ~2% to ~10% during the report period, peaking during the height of DeepSeek’s viral moment.
The below breaks out the message share among models with reasoning capabilities as a subcategory of text.
Usage of Gemini 2.5 Pro has been growing quickly among Poe subscribers, with the model obtaining a reasoning message share of ~30% within only ~6 weeks of its launch.
OpenAI, after releasing the category-defining reasoning model o1-preview in late 2024, continues to release more capable and affordable reasoning models at a pace unmatched by other labs, with the launch of o1-pro, o3-mini, o3-mini-high, o3, and o4-mini in just the first four months of 2025. Within the OpenAI set of reasoning models, it appears subscribers are quickly adopting the latest (e.g. o3-mini → o4-mini, o1 → o3) quite fluidly.
While xAI’s Grok 3 topped various problem-solving benchmarks in its February 2025 public release, Grok-3-mini continues to be the only model in the family that supports reasoning in the xAI API, which is perhaps why it accounts for less than 1% of reasoning model usage.
We note the early emergence of hybrid reasoning models, such as Gemini 2.5 Flash Preview and Qwen3, which can decide (or can be controlled) to vary their reasoning level conversationally (i.e., not just via API parameters). However, their collective usage remains small at ~1% in the subcategory.
Image generation is becoming increasingly competitive as quality and adherence improve
GPT Image Generation (GPT-Image-1) launched in the API in late April and rapidly attained 17% of image generation usage in only two weeks, mirroring its viral launch throughout March and early April in the ChatGPT app.
Google’s Imagen3 family has continued its steady usage growth from ~10% to ~30% share over the course of 2025, putting it on par with category-leader Black Forest Labs’ FLUX family of image generation models, which held ~35% share collectively as of last week of April.
The FLUX family of image generation models maintained its overall plurality share of image generation on Poe, but it declined slightly from ~45% to ~35% during the report period.
Kling 2.0 quickly emerged as a top contender in video generation in just three weeks. [4]
The newly released Kling family of video generation models from Chinese lab Kuaishou have rapidly garnered a collective ~30% usage share, most notably Kling-2.0-Master, which yielded 21% of all video generation on Poe at the end of April 2025, only three weeks after its release.
Google’s Veo 2 continues to maintain strong usage share at ~20% in the months following its February launch.
Category-defining video generation incumbent Runway has seen its usage share in video generation decline by ~40% to ~20% throughout the report period. [5]
ElevenLabs maintains its lead in audio generation amidst early signs of rising competition [6]
In audio generation (specifically text-to-speech, or “TTS”), ElevenLabs appears to be preferred by users with it fulfilling ~80% of all subscribers’ TTS requests during the report period.
However, competition is brewing in this space with the emergence of Cartesia, Unreal Speech, PlayAI, and Orpheus, which offer unique voice options, voice effects, and different performance and price profiles.
Conclusion
We hope sharing data from Poe's diverse user base and official integrations offers a valuable, real-world perspective on the dynamic and ever-evolving AI landscape. The increasing model diversity and provider competition helps underscore the value of our platform, for both users and creators. We’re currently observing break-out usage among reasoning models, and expect to see this continue as a top competitive driver among leading frontier labs. Multimedia is heating up and following OpenAI’s breakout offering in its new image generation capabilities, it may not be long before we see something similar among video models.
We look forward to continuing to share these important insights as we capture signs of new patterns and emerging trends. Finally, if you would like to experience access to our library of 100+ official model integrations, you can sign up on Poe today at https://poe.com/.
