State of AI

Our analysis is based on metadata collected from the OpenRouter platform, a unified AI inference layer that connects users and developers to hundreds of large language models. Each user request on OpenRouter is executed against a user-selected model, and structured metadata describing the resulting "generation" event is logged. The dataset used in this study consists of anonymized request-level metadata for billions of prompt–completion pairs from a global user base, spanning approximately two years up to the time of writing. We do zoom in on the last year.

Crucially, we did not have access to the underlying text of prompts or completions. Our analysis relies entirely on metadata that capture the structure, timing, and context of each generation, without exposing user content. This privacy-preserving design enables large-scale behavioral analysis.

Each generation record includes information on timing, model and provider identifiers, token usage, and system performance metrics. Token counts encompass both prompt (input) and completion (output) tokens, allowing us to measure overall model workload and cost. Metadata also include fields related to geographic routing, latency, and usage context (for example, whether the request was streamed or cancelled, or whether tool-calling features were invoked). Together, these attributes provide a detailed but non-textual view of how models are used in practice.

All analyses, aggregations, and most visualizations based on this metadata were conducted using the Hex analytics platform, which provided a reproducible pipeline for versioned SQL queries, transformations, and final figure generation.

We emphasize that this dataset is observational: it reflects real-world activity on the OpenRouter platform, which itself is shaped by model availability, pricing, and user preferences. As of 2025, OpenRouter supports more than 300+ active models from over 60 providers and serves millions of developers and end-users, with over 50% of usage originating outside the United States. While certain usage patterns outside the platform are not captured, OpenRouter's global scale and diversity make it a representative lens on large-scale LLM usage dynamics.

No direct access to user prompts or model outputs was available for this study. Instead, OpenRouter performs internal categorization on a random sample comprising approximately 0.25% of all prompts and responses through a non-proprietary module GoogleTagClassifier. While this represents only a fraction of total activity, the underlying dataset remains substantial given the overall query volume processed by OpenRouter. GoogleTagClassifier interfaces with Google Cloud Natural Language's classifyText content-classification API

The API applies a hierarchical, language-agnostic taxonomy to textual input, returning one or more category paths (e.g., /Computers & Electronics/Programming, /Arts & Entertainment/Roleplaying Games) with corresponding confidence scores in the range [0,1]. The classifier operates directly on prompt data (up to the first 1,000 characters). The classifier is deployed within OpenRouter's infrastructure, ensuring that classifications remain anonymous and are not linked to individual customers. Categories with confidence scores below the default threshold of 0.5 are excluded from further analysis. The classification system itself operates entirely within OpenRouter's infrastructure and was not part of this study; our analysis relied solely on the resulting categorical outputs (effectively metadata describing prompt classifications) rather than the underlying prompt content.

To make these fine-grained labels useful at scale, we map GoogleTagClassifier's taxonomy to a compact set of study-defined buckets and assign each request tags. Each tag rolls up to higher level category in one to one way. Representative mappings include:

• Programming: from /Computers & Electronics/Programming or /Science/Computer Science/*
• Roleplay: from /Games/Roleplaying Games and creative dialogue leaves under /Arts & Entertainment/*
• Translation: from /Reference/Language Resources/*
• General Q&A / Knowledge: from /Reference/General Reference/* and /News/* when the intent appears to be factual lookup
• Productivity/Writing: from /Computers & Electronics/Software/Business & Productivity Software or /Business & Industrial/Business Services/Writing & Editing Services
• Education: from /Jobs & Education/Education/*
• Literature/Creative Writing: from /Books & Literature/* and narrative leaves under /Arts & Entertainment/*
• Adult: from /Adult
• Others: for the long tail of prompts when no dominant mapping applies. (Note: we omit this category from most analyses below.)

There are inherent limitations to this approach, for instance, reliance on a predefined taxonomy constrains how novel or cross-domain behaviors are categorized, and certain interaction types may not yet fit neatly within existing classes. In practice, some prompts receive multiple category labels when their content spans overlapping domains. Nonetheless, the classifier-driven categorization provides us with a lens for downstream analyses. This enables us to quantify not just how much LLMs are used but what for.

A few variants are worth explicitly calling out:

• Open Source vs. Proprietary: We label models as open source (OSS, for simplicity) if their weights are publicly available, and closed source if access is only via a restricted API (e.g., Anthropic's Claude). This distinction lets us measure adoption of community-driven models versus proprietary ones.
• Origin (Chinese vs. Rest-of-World): Given the rise of Chinese LLMs and their distinct ecosystems, we tag models by primary locale of development. Chinese models include those developed by organizations in China, Taiwan, or Hong Kong (e.g., Alibaba's Qwen, Moonshot AI's Kimi, or DeepSeek). RoW (Rest-of-World) models cover North America, Europe, and other regions.
• Prompt vs. Completion Tokens: We distinguish between prompt tokens, which represent the input text provided to a model, and completion tokens, which represent the model's generated output. Total tokens equal the sum of prompt and completion tokens. Reasoning tokens represent internal reasoning steps in models with native reasoning capabilities and are included within completion tokens.

Unless otherwise noted, token volume refers to the sum of prompt (input) and completion (output) tokens.

To understand regional patterns in LLM usage, we segment requests by user geography. Direct request metadata (like IP-based location) is typically imprecise or anonymized. Instead, we determine user region based on the billing location associated with each account. This provides a more reliable proxy for user geography, as billing data reflects the country or region linked to the user's payment method or account registration. We use this billing-based segmentation in our analysis of regional adoption and model preferences.

This method has limitations. Some users employ third-party billing or shared organizational accounts, which may not correspond to their actual location. Enterprise accounts may aggregate activity across multiple regions under one billing entity. Despite these imperfections, billing geography remains the most stable and interpretable indicator available for privacy-preserving geographic analysis given the metadata we had access to.

Our analyses primarily cover a rolling 13-month period ending on November, 2025, but not all underlying metadata spans this full window. Most model-level and pricing analyses were focused on November 3, 2024 – November 30, 2025 time frame. However, category-level analyses (especially those using the GoogleTagClassifier taxonomy) are based on a shorter interval beginning in May 2025, reflecting when consistent tagging became available on OpenRouter. In particular, detailed task classification fields (e.g., tags such as Programming, Roleplay, or Technology) were only added in mid-2025. Consequently, all findings in the Categories section should be interpreted as representative of mid-2025 usage rather than the entire prior year.

Unless otherwise specified, all time-series aggregates are computed on a weekly basis using UTC-normalized timestamps, summing prompt and completion tokens. This approach ensures comparability across model families and minimizes bias from transient spikes or regional time-zone effects.

The table below ranks the top model families in our dataset by total token volume served. The landscape of OSS models has shifted significantly over the last year: while DeepSeek remains the single largest OSS contributor by volume, its dominance has waned as new entrants rapidly gain ground. Today, multiple open source families each sustain substantial usage, pointing to a diversified ecosystem.

This figure illustrates the dramatic evolution of market share among the top individual open source models week by week. Early in the period (late 2024), the market was highly consolidated: two models from the DeepSeek family (V3 and R1) consistently accounted for over half of all OSS token usage, forming the large, dark blue bands at the bottom of the chart.

This near-monopoly structure shattered following the Summer Inflection (mid-2025). The market has since become both broader and deeper, with usage diversifying significantly. New entrants like Qwen's models, Minimax's M2, MoonshotAI's Kimi K2, and OpenAI's GPT-OSS series all grew rapidly to serve significant portions of requests, often achieving production-scale adoption within weeks of release. This signals that the open source community and AI startups can achieve quick adoption by introducing models with novel capabilities or superior efficiency.

By late 2025, the competitive balance had shifted from near-monopoly to a pluralistic mix. No single model exceeds 25% of OSS tokens, and the token share is now distributed more evenly across five to seven models. The practical implication is that users are finding value in a wider array of options, rather than defaulting to one "best" choice. Although this figure visualizes relative share among OSS models (not absolute volume), the clear trend is a decisive shift toward market fragmentation and increased competition within the open source ecosystem.

Overall, the open source model ecosystem is now highly dynamic. Key insights include:

• Top-tier diversity: Where one family (DeepSeek) once dominated OSS usage, we now increasingly see half a dozen models each sustaining meaningful share. No single open model holds more than ≈20–25% of OSS tokens consistently.
• Rapid scaling of new entrants: Capable new open models can capture significant usage within weeks. For example, MoonshotAI's models quickly grew to rival older OSS leaders, and even a newcomer like MiniMax went from zero to substantial traffic in a single quarter. This indicates low switching friction and a user base eager to experiment.
• Iterative advantage: The longevity of DeepSeek's presence at the top underscores that continuous improvement is critical. DeepSeek's successive releases (Chat-V3, R1, etc.) kept it competitive even as challengers emerged. OSS models that stagnate in development tend to lose share to those with frequent updates at the frontier or domain-specific finetunes.

Today the open source LLM arena in 2025 resembles a competitive ecosystem where innovation cycles are rapid and leadership is not guaranteed. For model builders, this means that releasing an open model with state-of-the-art performance can yield immediate uptake, but maintaining usage share requires ongoing investment in further development. For users and application developers, the trend is positive: there is a richer selection of open models to choose from, often with comparable or sometimes superior capabilities to proprietary systems in specific areas (like roleplay).

A year ago, the open source model ecosystem was largely a story of trade-offs between two extremes: a vast number of small, fast models and a handful of powerful, large-scale models. However, a review of the past year reveals a significant maturation of the market and the emergence of a new, growing category: the medium-sized model. Please note that we categorize models by their parameter count as follows:

• Small: Models with fewer than 15 billion parameters.
• Medium: Models with 15 billion to 70 billion parameters.
• Large: Models with 70 billion or more parameters.

The data on developer and user behavior tells us a nuanced story. The figures show that while the number of models across all categories has grown, the usage has shifted notably. Small models are losing favor while medium and large models are capturing that value.

A deeper look at the models driving these trends reveals distinct market dynamics:

• The "Small" Market: Overall Decline in Usage. Despite a steady supply of new models, the small model category as a whole is seeing its share of usage decline. This category is characterized by high fragmentation. No single model holds a dominant position for long, and it sees a constant churn of new entrants from a diverse set of providers like Meta, Google, Mistral, and DeepSeek. For example, Google Gemma 3.12B (released August 2025) saw a rapid adoption but competes in a crowded field where users continually seek the next best alternative.
• The "Medium" Market: Finding "Model-Market Fit." The medium model category tells a clear story of market creation. The segment itself was negligible until the release of Qwen2.5 Coder 32B in November 2024, which effectively established this category. This segment then matured into a competitive ecosystem with the arrival of other strong contenders like Mistral Small 3 (January 2025) and GPT-OSS 20B (August 2025), which carved out user mind share. This segment demonstrates that users are seeking a balance of capability and efficiency.
• The "Large" Model Segment: A Pluralistic Landscape. The "flight to quality" has not led to consolidation but to diversification. The large model category now features a range of high-performing contenders from Qwen3 235B A22B Instruct (released in July 2025) and Z.AI GLM 4.5 Air to OpenAI: GPT-OSS-120B (August 5th): each capturing meaningful and sustained usage. This pluralism suggests users are actively benchmarking across multiple open large models rather than converging on a single standard.

The era of small models dominating the open source ecosystem might be behind. The market is now bifurcating, with users either gravitating toward a new, robust class of medium models, or consolidating their workloads onto the single most capable large model.

Open-source models today are employed for a remarkably broad range of tasks, spanning creative, technical, and informational domains. While proprietary models still dominate in structured business tasks, OSS models have carved out leadership in two particular areas: creative roleplay and programming assistance. Together, these categories account for the majority of OSS token usage.

The figure above highlights that more than half of all OSS model usage falls under Roleplay, with Programming being the second-largest category. This indicates that users turn to open models primarily for creative interactive dialogues (such as storytelling, character roleplay, and gaming scenarios) and for coding-related tasks. The dominance of roleplay (hovering at more than 50% of all OSS tokens) underscores a use case where open models have an edge: they can be utilized for creativity and are often less constrained by content filters, making them attractive for fantasy or entertainment applications. Roleplay tasks require flexible responses, context retention, and emotional nuance - attributes that open models can deliver effectively without being heavily restricted by commercial safety or moderation layers. This makes them particularly appealing for communities experimenting with character-driven experiences, fan fiction, interactive games, and simulation environments.

The figure above shows category breakdown over time if we zoom in on Chinese OSS models only. These models are no longer used primarily for creative tasks. Roleplay remains the largest category at around 33%, but programming and technology now account for a combined majority of usage (39%). This shift suggests that models like Qwen and DeepSeek are increasingly used for code generation and infrastructure-related workloads. While high-volume enterprise users may influence specific segments, the overall trend points to Chinese OSS models competing directly in technical and productivity domains.

If we zoom in just on the programming category, we observe that proprietary models still handle the bulk of coding assistance overall (the gray region), reflecting strong offerings like Anthropic's Claude. However, within the OSS portion, there was a notable transition: in mid-2025, Chinese OSS models (blue) delivered the majority of open source coding help (driven by early successes like Qwen 3 Coder). By Q4 2025, Western OSS models (orange) such as Meta's LLaMA-2 Code and OpenAI's GPT-OSS series had surged, but decreased in overall share in recent weeks. This oscillation suggests a very competitive environment. The practical takeaway is that open source code assistant usage is dynamic and highly responsive to new model quality: developers are open to whichever OSS model currently provides the best coding support. As a limitation, this figure doesn't show absolute volumes: open source coding usage grew overall so a shrinking blue band doesn't mean Chinese OSS lost users, only relative share.

Now if we examine just the roleplay traffic, we see that it is now almost equally served by Rest-of-World OSS (orange, 43% in recent weeks) and Closed (gray, at ~42% most recently) models. This represents a significant shift from earlier in 2025, when the category was dominated by proprietary (gray) models, which held approximately 70% of the token share. At that time (May 2025), Western OSS models accounted for only ~22% of traffic, and Chinese OSS (blue) models held a small share of ~8%. Throughout the year, the proprietary share steadily eroded. By the end of October 2025, this trend accelerated as both Western and Chinese open source models gained significant ground.

The resulting convergence indicates a healthy competition; users have viable choices from both open and proprietary offerings for creative chats and storytelling. This reflects that developers recognize the demand for roleplay/chat models and have tailored their releases to that end (e.g., fine-tuning on dialogues, adding alignment for character consistency). A point to note is that "roleplay" covers a range of subgenres (from casual chatting to complex game scenarios). Yet from a macro perspective, it is clear OSS models have an edge in this creative arena.

Interpretation. Broadly, across the OSS ecosystem, the key use cases are: Roleplay and creative dialogue: the top category, likely because open models can be uncensored or more easily customized for fictional persona and story tasks. Programming assistance: second-largest, and growing, as open models become more competent at code. Many developers leverage OSS models locally for coding to avoid API costs. Translation and multilingual support: a steady use case, especially with strong bilingual models available (Chinese OSS models have an edge here). General knowledge Q&A and education: moderate usage; while open models can answer questions, users may prefer closed models like GPT-5 for highest factual accuracy.

It is worth noting that the OSS usage pattern (heavy on roleplay) mirrors what many might consider for "enthusiasts" or "indie developers" - areas where customization and cost-efficiency trump absolute accuracy. The lines are blurring, though: OSS models are rapidly improving in technical domains, and proprietary models are being used creatively too.

As shown in the figure above, the share of total tokens routed through reasoning-optimized models climbed sharply in 2025. What was effectively a negligible slice of usage in early Q1 now exceeds fifty percent. This shift reflects both sides of the market. On the supply side, the release of higher-capability systems like GPT-5, Claude 4.5, and Gemini 3 expanded what users could expect from stepwise reasoning. On the demand side, users increasingly prefer models that can manage task state, follow multi-step logic, and support agent-style workflows rather than simply generate text.

The figure above shows the top models driving this shift. In the most recent data, xAI's Grok Code Fast 1 now drives the largest share of reasoning traffic (excluding free launch access), ahead of Google's Gemini 2.5 Pro and Gemini 2.5 Flash. This is a notable change from only a few weeks ago, when Gemini 2.5 Pro led the category and DeepSeek R1 and Qwen3 were also in the top tier. Grok Code Fast 1 and Grok 4 Fast have gained share quickly, supported by xAI's aggressive rollout, competitive pricing, and developer attention around its code-oriented variants. At the same time, the continued presence of open models like OpenAI's gpt-oss-120b underscores that developers still reach for OSS when possible. The mix overall highlights how dynamic the reasoning landscape has become, with rapid model turnover shaping which systems dominate real workloads.

The data points to a clear conclusion: reasoning-oriented models are becoming the default path for real workloads, and the share of tokens flowing through them is now a leading indicator of how users want to interact with AI systems.

In the figure above, we report the share of total tokens originating from requests whose finish reason was a Tool Call. This metric is normalized and captures only those interactions in which a tool was actually invoked.

This is in contrast to the Input Tool signal that records whether a tool was provided to the model during a request (regardless of invocation). Input Tool counts are, by definition, higher than Tool Call finish reasons, since provision is a superset of successful execution. Whereas the finish-reason metric measures realized tool use, Input Tool reflects potential availability rather than actual invocation. Because this metric was introduced only in September 2025, we are not reporting it in this paper.

The noticeable spike in May in the figure above was largely attributable to one sizable account whose activity briefly lifted overall volumes. Aside from this anomaly, tool adoption has shown a consistent upward trend throughout the year.

As shown in the figure above, tool invocation was initially concentrated among a small group of models: OpenAI's gpt-4o-mini and Anthropic's Claude 3.5 and 3.7 series, which together accounted for most tool-enabled tokens in early 2025. By mid-year, however, a broader set of models began supporting tool provision, reflecting a more competitive and diversified ecosystem. From end of September onward, newer Claude 4.5 Sonnet model rapidly gained share. Meanwhile, newer entries like Grok Code Fast and GLM 4.5 have made visible inroads, reflecting broader experimentation and diversification in tool-capable deployments.

For operators, the implication is clear: enabling tool use is on the rise for high-value workflows. Models without reliable tool formats risk falling behind in enterprise adoption and orchestration environments.

The shape of model workloads has evolved markedly over the past year. Both prompt (input) and completion (output) token volumes have risen sharply, though at different scales and rates. Average prompt tokens per request have increased roughly fourfold from around 1.5K to over 6K while completions have nearly tripled from about 150 to 400 tokens. The relative magnitude of growth highlights a decisive shift toward more complex, context-rich workloads.

This pattern reflects a new equilibrium in model usage. The typical request today is less about open-ended generation ("write me an essay") and more about reasoning over substantial user-provided material such as codebases, documents, transcripts, or long conversations, and producing concise, high-value insights. Models are increasingly acting as analytical engines rather than creative generators.

Category-level data (available only since Spring 2025) provides a more nuanced picture: programming workloads are the dominant driver of prompt token growth. Requests involving code understanding, debugging, and code generation routinely exceed 20K input tokens, while all other categories remain relatively flat and low-volume. This asymmetric contribution suggests that the recent expansion in prompt size is not a uniform trend across tasks but rather a concentrated surge tied to software development and technical reasoning use cases.

Sequence length is a proxy for task complexity and interaction depth. The figure above shows that average sequence length has more than tripled over the past 20 months from under 2,000 tokens in late 2023 to over 5,400 by late 2025. This growth reflects a structural shift toward longer context windows, deeper task history, and more elaborate completions.

As per previous section, the second figure adds further clarity: programming-related prompts now average 3–4 times the token length of general-purpose prompts. The divergence indicates that software development workflows are the primary driver of longer interactions. Long sequences are not just user verbosity: they are a signature of embedded, more sophisticated agentic workflows.

Together, these trends (rising reasoning share, expanded tool use, longer sequences, and programming's outsize complexity) suggest that the center of gravity in LLM usage has shifted. The median LLM request is no longer a simple question or isolated instruction. Instead, it is part of a structured, agent-like loop, invoking external tools, reasoning over state, and persisting across longer contexts.

For model providers, this raises the bar for default capabilities. Latency, tool handling, context support, and robustness to malformed or adversarial tool chains are increasingly critical. For infra operators, inference platforms must now manage not just stateless requests but long-running conversations, execution traces, and permission-sensitive tool integrations. Soon enough, if not already, agentic inference will be taking over the majority of the inference.

Programming has become the most consistently expanding category across all models. The share of programming-related requests has grown steadily through 2025, paralleling the rise of LLM-assisted development environments and tool integrations. As shown in the figure above, programming queries accounted for roughly 11% of total token volume in early 2025 and exceeded 50% in recent weeks. This trend reflects a shift from exploratory or conversational use toward applied tasks such as code generation, debugging, and data scripting. As LLMs become embedded in developer workflows, their role as programming tools is being normalized. This evolution has implications for model development, including increased emphasis on code-centric training data, improved reasoning depth for multi-step programming tasks, and tighter feedback loops between models and integrated development environments.

This growing demand for programming support is reshaping competitive dynamics across model providers. As shown in the figure below, Anthropic's Claude series has consistently dominated the category, accounting for more than 60% of programming-related spend for most of the observed period. The landscape has nevertheless evolved meaningfully. During the week of November 17, Anthropic's share fell below the 60% threshold for the first time. Since July, OpenAI has expanded its share from roughly 2% to about 8% in recent weeks, likely reflecting a renewed emphasis on developer-centric workloads. Over the same interval, Google's share has remained stable at approximately 15%. The mid-tier segment is also in motion. Open source providers including Z.AI, Qwen, and Mistral AI are steadily gaining mindshare. MiniMax, in particular, has emerged as a fast-rising entrant, showing notable gains in recent weeks.

Overall, programming has become one of the most contested and strategically important model categories. It attracts sustained attention from top labs, and even modest changes in model quality or latency can shift share week to week. For infrastructure providers and developers, this highlights the need for continual benchmarking and evals, especially as the frontier is constantly evolving.

The figures above break down LLM usage across the twelve most common content categories, revealing the internal sub-topic structure of each. A key takeaway is that most categories are not evenly distributed: they are dominated by one or two recurring use patterns, often reflecting concentrated user intent or alignment with LLM strengths.

Among the highest-volume categories, roleplay stands out for its consistency and specialization. Nearly 60% of roleplay tokens fall under Games/Roleplaying Games, suggesting that users treat LLMs less as casual chatbots and more as structured roleplaying or character engines. This is further reinforced by the presence of Writers Resources (15.6%) and Adult content (15.4%), pointing to a blend of interactive fiction, scenario generation, and personal fantasy. Contrary to assumptions that roleplay is mostly informal dialogue, the data show a well-defined and replicable genre-based use case.

Programming is similarly skewed, with over two-thirds of traffic labeled as Programming/Other. This signals the broad and general-purpose nature of code-related prompts: users are not narrowly focused on specific tools or languages but are asking LLMs for everything from logic debugging to script drafting. That said, Development Tools (26.4%) and small shares from scripting languages indicate emerging specialization. This fragmentation highlights an opportunity for model builders to improve tagging or training around structured programming workflows.

Beyond the dominant categories of roleplay and programming, the remaining domains represent a diverse but lower-volume tail of LLM usage. While individually smaller, they reveal important patterns about how users interact with models across specialized and emerging tasks. For example, translation, science, and health show relatively flat internal structure. In translation, usage is nearly evenly split between Foreign Language Resources (51.1%) and Other, suggesting diffuse needs: multilingual lookup, rephrasing, light code-switching, rather than sustained document-level translation. Science is dominated by a single tag, Machine Learning & AI (80.4%), indicating that most scientific queries are meta-AI questions rather than general STEM topics like physics or biology. This reflects either user interest or model strengths skewed toward self-referential inquiry.

Health, in contrast, is the most fragmented of the top categories, with no sub-tag exceeding 25%. Tokens are spread across medical research, counseling services, treatment guidance, and diagnostic lookups. This diversity highlights the domain's complexity, but also the challenge of modeling it safely: LLMs must span high variance user intent, often in sensitive contexts, without clear concentration in a single use case.

What links these long-tail categories is their broadness: users turn to LLMs for exploratory, lightly structured, or assistance-seeking interactions, but without the focused workflows seen in programming or personal assistants. Taken together, these secondary categories may not dominate volume, but they hint at latent demand. They signal that LLMs are being used at the fringes of many fields from translation to medical guidance to AI introspection and that as models improve in domain robustness and tooling integration, we may see these scattered intents converge into clearer, higher-volume applications.

By contrast, finance, academia, and legal are much more diffuse. Finance spreads its volume across foreign exchange, socially responsible investing, and audit/accounting: no single tag breaks 20%. Legal shows similar entropy, with usage split between Government/Other (43.0%) and Legal/Other (17.8%). This fragmentation may reflect the complexity of these domains, or simply the lack of targeted LLM workflows for them compared to more mature categories like coding and chat.

The data suggest that real-world LLM usage is not uniformly exploratory: it clusters tightly around a small set of repeatable, high-volume tasks. Roleplay, programming, and personal assistance each exhibit clear structure and dominant tags. Science, health, and legal domains, by contrast, are more diffuse and likely under-optimized. These internal distributions can guide model design, domain-specific fine-tuning, and application-level interfaces particularly in tailoring LLMs to user goals.

Different model authors are utilized in different usage patterns. The figures below show the distribution of content categories for major model families (Anthropic's Claude, Google's models, OpenAI's GPT series, DeepSeek, and Qwen). Each bar represents 100% of that provider's token usage, broken down by top tags.

Anthropic's Claude is heavily skewed toward Programming + Technology uses, which together exceed 80% of its usage. Roleplay and general Q&A are only a small sliver. This confirms Claude's positioning as a model optimized for complex reasoning, coding, and structured tasks; developers and enterprises appear to use Claude mainly as a coding assistant and problem solver.

Google's model usage is more diverse. We see notable segments for Translation, Science, Technology, and some General Knowledge. For instance, ~5% of Google's usage was legal or policy content, and another ~10% science-related. It might hint at Gemini's broad training focus. Compared to others, Google has relatively less and, in fact, declining coding share by late 2025 (down to roughly 18%) and a broader tail of categories. This suggests Google's models are being used more as general-purpose information engines.

xAI's usage profile is distinct from the other providers. For most of the period, usage is overwhelmingly concentrated in Programming, often exceeding eighty percent of all tokens. Only in late November does the distribution broaden, with noticeable gains in Technology, Roleplay, and Academia. This sharp shift aligns with the timing of xAI's model being distributed at no cost through select consumer applications, which likely introduced a large influx of non-developer traffic. The result is a usage composition that blends an early, developer-heavy core with a sudden wave of general-purpose engagement, suggesting that xAI's adoption path is being shaped both by technical users and by episodic surges tied to promotional availability.

OpenAI's usage profile has shifted markedly through 2025. Earlier in the year, science tasks accounted for more than half of all OpenAI tokens; by late 2025, that share had declined to under 15%. Meanwhile, programming and technology-related usage now comprise more than half of total volume (29% each), reflecting deeper integration into developer workflows, productivity tools, and professional applications. OpenAI's usage composition now sits between Anthropic's tightly focused profile and Google's more diffuse distribution, suggesting a broad base of utility with growing tilt toward high-value, structured tasks.

DeepSeek and Qwen exhibit usage patterns that diverge considerably from the other model families discussed earlier. DeepSeek's token distribution is dominated by roleplay, casual chat, and entertainment-oriented interaction, often accounting for more than two thirds of its total usage. Only a small fraction of activity falls into structured tasks such as programming or science. This pattern reflects DeepSeek's strong consumer orientation and its positioning as a high-engagement conversational model. Notably, DeepSeek displays a modest but steady increase in programming-related usage toward late summer, suggesting incremental adoption in lightweight development workflows.

Qwen, by contrast, presents an almost inverted profile. Across the entire period shown, programming consistently represents 40-60 percent of all tokens, signaling a clear emphasis on technical and developer tasks. Compared with Anthropic's more stable engineering-heavy composition, Qwen demonstrates higher volatility across adjacent categories such as science, technology, and roleplay. These week-to-week shifts imply a heterogeneous user base and rapid iteration in applied use cases. A noticeable rise in roleplay usage during September and October, followed by a contraction in November, hints at evolving user behavior or adjustments in downstream application routing.

In summary, each provider shows a distinct profile aligned with its strategic focus. The differences highlight why no single model or provider covers all use cases optimally; it also underscores the potential benefits of multi-model ecosystem.

The distribution of spend, as shown in the figure below, underscores the increasingly global nature of AI inference market. North America, while still the single largest region, now accounts for less than half of total spend for most of the observed period. Europe shows a stable and durable contribution. Its relative share of weekly spend remains consistent throughout the timeline, typically occupying a band between the mid-teens and low twenties. A notable development is the rise of Asia not only as a producer of frontier models but also as a rapidly expanding consumer. In the earliest weeks of the dataset, Asia represented roughly thirteen percent of global spend. Over time, this share more than doubled, reaching approximately 31% in the most recent period.

As shown in the table above, English dominates usage, accounting for more than 80% of all tokens. This reflects both the prevalence of English-language models and the developer-centric skew of OpenRouter's user base. However, other languages particularly Chinese, Russian, and Spanish, make up a meaningful tail. Simplified Chinese alone accounts for nearly 5% of global tokens, suggesting sustained engagement by users in bilingual or Chinese-first environments, especially given the growth of Chinese OSS models like DeepSeek and Qwen.

Cohort Retention Rates. Retention is measured as activity retention, where users are counted if they return in subsequent months, even after periods of inactivity; as a result, curves may exhibit small non-monotonic bumps.

This collection of retention charts captures the dynamics of the LLM user market across leading models. At first glance, the data is dominated by high churn and rapid cohort decay. Yet beneath this volatility lies a subtler and more consequential signal: a small set of early user cohorts exhibits durable retention over time. We term these foundational cohorts.

These cohorts are not merely early adopters; they represent users whose workloads have achieved a deep and persistent workload–model fit. Once established, this fit creates both economic and cognitive inertia that resists substitution, even as newer models emerge.

We introduce the Cinderella Glass Slipper effect as a framework to describe this phenomenon. The hypothesis posits that in a rapidly evolving AI ecosystem, there exists a latent distribution of high-value workloads that remain unsolved across successive model generations. Each new frontier model is effectively "tried on" against these open problems. When a newly released model happens to match a previously unmet technical and economic constraint, it achieves the precise fit — the metaphorical "glass slipper."

For the developers or organizations whose workloads finally "fit," this alignment creates strong lock-in effects. Their systems, data pipelines, and user experiences become anchored to the model that solved their problem first. As costs decline and reliability increases, the incentive to re-platform diminishes sharply. Conversely, workloads that do not find such a fit remain exploratory, migrating from one model to another in search of their own solution.

Empirically, this pattern is observable in the June 2025 cohort of Gemini 2.5 Pro and the May 2025 cohort of Claude 4 Sonnet, which retain approximately 40% of users at Month 5, substantially higher than later cohorts. These cohorts appear to correspond to specific technical breakthroughs (e.g., reasoning fidelity or tool-use stability) that finally enabled previously impossible workloads.

• First-to-Solve as Durable Advantage. The classical first-mover advantage gains significance when a model is the first to solve a critical workload. Early adopters embed the model across pipelines, infrastructure, and user behaviors, resulting in high switching friction. This creates a stable equilibrium in which the model retains its foundational cohort even as newer alternatives emerge.
• Retention as an Indicator of Capability Inflection. Cohort-level retention patterns serve as empirical signals of model differentiation. Persistent retention in one or more early cohorts indicates a meaningful capability inflection — a workload class that transitions from infeasible to possible. Absence of such patterns suggests capability parity and limited depth of differentiation.
• Temporal Constraints of the Frontier Window. The competitive landscape imposes a narrow temporal window in which a model can capture foundational users. As successive models close the capability gap, the probability of forming new foundational cohorts declines sharply. The "Cinderella" moment, where model and workload align precisely, is thus transient but decisive for long-term adoption dynamics.

In all, rapid capability shifts in foundation models necessitate a redefinition of user retention. Each new model generation introduces a brief opportunity to solve previously unmet workloads. When such alignment occurs, the affected users form foundational cohorts: segments whose retention trajectories remain stable despite subsequent model introductions.

The Dominant Launch Anomaly. The OpenAI GPT-4o Mini chart shows this phenomenon in its extreme. A single foundational cohort (July 2024, orange line) established a dominant, sticky workload-model fit at launch. All subsequent cohorts, which arrived after this fit was established and the market had moved on, behave identically: they churn and cluster at the bottom. This suggests the window to establish this foundational fit is singular and occurs only at the moment a model is perceived as "frontier."

The Consequence of No-Fit. The Gemini 2.0 Flash and Llama 4 Maverick charts showcase a cautionary tale of what happens when this initial fit is never established. Unlike the other models, there is no high-performing foundational cohort. Every single cohort performs identically poorly. This suggests that the models were never perceived as a "frontier" for a high-value, sticky workload. It launched directly into the good enough market and thus failed to lock in any user base. Similarly, the chaotic charts for DeepSeek, despite overwhelming success overall, struggle to establish a stable, foundational cohort.

Boomerang Effect. The DeepSeek models introduce a more complex pattern. Their retention curves display a highly unusual anomaly: resurrection jumps. Unlike typical, monotonically decreasing retention, several DeepSeek cohorts show a distinct rise in retention after an initial period of churn (e.g., DeepSeek R1's April 2025 cohort around Month 3, and DeepSeek Chat V3-0324's July 2025 cohort around Month 2). This indicates that some churned users are returning to the model. This "boomerang effect" suggests these users return to DeepSeek, after trying alternatives and confirming through competitive testing that DeepSeek provides an optimal, and often better fit for their specific workload due to a superior combination of specialized technical performance, cost-efficiency, or other unique features.

Implications. The Glass Slipper phenomenon reframes retention not as an outcome but as a lens for understanding capability breakthroughs. Foundational cohorts are the fingerprints of real technical progress: they mark where an AI model has crossed from novelty into necessity. For builders and investors alike, identifying these cohorts early may be the single most predictive signal of enduring model–market advantage.

The scatter plot above reveals a distinct segmentation of AI use cases, mapping them based on their aggregate usage volume (Total Tokens) against their unit cost (Cost per 1M Tokens). A critical preliminary observation is that both axes are logarithmic. This logarithmic scaling signifies that small visual distances on the chart correspond to substantial multiplicative differences in real-world volume and cost.

The chart is bisected by a vertical line at the median cost of $0.73 per 1M tokens, effectively creating a four-quadrant framework to simplify the AI market across categories.

Note that these end costs differ from advertised list prices. High-frequency workloads benefit from caching, which drives down realized spend and produces materially lower effective prices than those publicly listed. The cost metric shown reflects a blended rate across both prompt and completion tokens, providing a more accurate view of what users actually pay in aggregate. The dataset also excludes BYOK activity to isolate standardized, platform-mediated usage and avoid distortion from custom infrastructure setups.

Premium Workloads (Top-Right): This quadrant contains high-cost, high-usage applications, now including technology and science, positioned right at the intersection. These represent valuable and heavily-used professional workloads where users are willing to pay a premium for performance or specialized capabilities. Technology is a significant outlier, being dramatically more expensive than any other category. This suggests that technology as a use case (perhaps relating to complex system design or architecture) may require far more powerful and expensive models for inference, yet it maintains a high usage volume, indicating its essential nature.

Mass-Market Volume Drivers (Top-Left): This quadrant is defined by high usage and a low, at-or-below-average cost. This area is dominated by two massive use cases: roleplay, programming as well as science.

• Programming stands out as the "killer professional" category, demonstrating the highest usage volume while having a highly optimized, median cost.
• Roleplay's usage volume is immense, nearly rivaling programming. This is a striking insight: a consumer-facing roleplay application drives a volume of engagement on par with a top-tier professional one.

The sheer scale of these two categories confirms that both professional productivity and conversational entertainment are primary, massive drivers for AI. The cost sensitivity in this quadrant is where, as previously noted, open source models have found a significant edge.

Specialized Experts (Bottom-Right): This quadrant houses lower-volume, high-cost applications, including finance, academia, health, and marketing. These are high-stakes, niche professional domains. The lower aggregate volume is logical, as one might consult an AI for "health" or "finance" far less frequently than for "programming." Users are willing to pay a significant premium for these tasks, likely because the demand for accuracy, reliability, and domain-specific knowledge is extremely high.

Niche Utilities (Bottom-Left): This quadrant features low-cost, low-volume tasks, including translation, legal, and trivia. These are functional, cost-optimized utilities. Translation has the highest volume within this group, while trivia has the lowest. Their low cost and relatively low volume suggest these tasks may be highly optimized, "solved," or commoditized, where good-enough alternative is available cheaply.

As noted, the most significant outlier on this chart is technology. It commands the highest cost-per-token by a substantial margin while maintaining high usage. This strongly suggests a market segment with a high willingness-to-pay for high-value, complex answers (e.g., system architecture, advanced technical problem-solving). One key question is whether this high price is driven by high user value (a "demand-side" opportunity) or by a high cost-of-serving (a "supply-side" challenge), as these queries may require the most powerful frontier models. The "play" to be had in technology is to service this high-value market. A provider who can serve this segment, perhaps through highly, optimized, specialist models, could potentially capture a market with higher margins.

The figure above maps model usage against cost per 1M tokens (log–log scale), revealing weak overall correlation. The x-axis maps out the nominal values for convenience. The trendline is nearly flat, indicating that demand is relatively price-inelastic; a 10% decrease in price corresponds to only about a 0.5–0.7% increase in usage. Yet the dispersion across the chart is substantial, reflecting strong market segmentation. Two distinct regimes appear: proprietary models from OpenAI and Anthropic occupy the high-cost, high-usage zone, while open models like DeepSeek, Mistral, and Qwen populate the low-cost, high-volume zone. This pattern supports a simple heuristic: closed source models capture high value tasks, while open source models capture high volume lower value tasks. The weak price elasticity indicates that even drastic cost differences do not fully shift demand; proprietary providers retain pricing power for mission-critical applications, while open ecosystems absorb volume from cost-sensitive users.

The figure above is similar to the prior figure but displays the model authors. Four usage–cost archetypes emerge. Premium leaders, such as Anthropic's Claude 3.7 Sonnet and Claude Sonnet 4, command costs around $2 per 1M tokens and still reach high usage, suggesting users are willing to pay for superior reasoning and reliability at scale. Efficient giants, like Google's Gemini 2.0 Flash and DeepSeek V3 0324, pair strong performance with prices below $0.40 per 1M tokens and achieve similar usage levels, making them attractive defaults for high-volume or long-context workloads. Long tail models, including Qwen 2 7B Instruct and IBM Granite 4.0 Micro, are priced at just a few cents per 1M tokens yet sit around 10^2.9 in total usage, reflecting constraints from weaker performance, limited visibility, or fewer integrations. Finally, premium specialists, such as OpenAI's GPT-4 and GPT-5 Pro, occupy the high-cost, low-usage quadrant: at roughly $35 per 1M tokens and usage near 10^3.4, they are used sparingly for niche, high-stakes workloads where output quality matters far more than marginal token cost.

Overall, the scatterplot highlights that pricing power in the LLM market is not uniform. While cheaper models can drive scale through efficiency and integration, premium offerings still command strong demand where stakes are high. This fragmentation suggests that the market has not yet commoditized, and that differentiation, whether through latency, context length, or output quality, remains a source of strategic advantage.

These observations suggest the following:

• At a macro level, demand is inelastic, but this masks different micro behaviors. Enterprises with mission-critical tasks will pay high prices (so these models see high usage). On the other hand, hobbyists and dev pipelines are very cost-sensitive and flock to cheaper models (leading to large usage for efficient models).
• There is some evidence of Jevons Paradox: making some models very cheap (and fast) led to people using them for far more tasks, ultimately consuming more total tokens. We see this in the efficient giants group: as cost per token dropped, those models got integrated everywhere and total consumption soared (people run longer contexts, more iterations, etc.).
• Quality and capabilities often trump cost: The heavy usage of expensive models (Claude, GPT-4) indicates that if a model is significantly better or has a trust advantage, users will bear higher costs. Often these models are integrated in workflows where the cost is negligible relative to the value of what they produce (e.g., code that saves an hour of developer time is worth far more than a few dollars of API calls).
• Conversely, simply being cheap isn't enough, a model must also be differentiable and sufficiently capable. Many open models priced near zero still because they are just good enough but don't find a workload-model fit or are not quite reliable, so developers hesitate to integrate them deeply.

From an operator's standpoint, several strategic patterns emerge. Providers like Google have leaned heavily into tiered offerings (most notably with Gemini Flash and Pro) explicitly trading off speed, cost, and capability. This tiering enables market segmentation by price sensitivity and task criticality: lightweight tasks are routed to cheaper, faster models; premium models serve complex or latency-tolerant workloads. Optimizing for use cases and reliability is often as impactful as "cutting" price. A faster, purpose-built model may be preferred over a cheaper but unpredictable one, especially in production settings. This shifts focus from cost-per-token to cost-per-successful-outcome. The relatively flat demand elasticity suggests LLMs are not yet a commodity—many users are willing to pay a premium for quality, capabilities, or stability. Differentiation still holds value, particularly when task outcomes matter more than marginal token savings.

The figures below break down the internal sub-tag structure for the three major domains: roleplay, programming, and technology. Each domain exhibits distinct internal patterns that reveal how users interact with LLMs within these categories.

All three domains (roleplay, technology, programming) exhibit distinct internal patterns, reflecting how users engage with LLMs across different sub-categories within each major domain.