XNews

Rise of Agentic and Reasoning AI

Rise of Agentic and Reasoning AI

OpenRouter and a16z released 'State of AI 2025' highlighting a shift to agentic reasoning models. - OpenRouter dataset: 100 trillion tokens of real interaction data - Report labels 2025 as AI's 'coming of age' - Reasoning models (e.g., OpenAI o1) enable planning, tool use, multi-step solutions - Reasoning models now account for over 50% of traffic - Prompt lengths increased roughly 4x - Tool invocation (search, code execution) has become routine - Open-source ecosystem diversified; mid-size (15B–70B) models favored - Chinese open models (DeepSeek, Qwen, Moonshot, ZAI) reached ~30% of open-source share - Top real user use cases: programming (fastest growth) and roleplay (~50% of single-app usage) - NeurIPS-winning paper: gating in attention improves long-context handling and training stability - OpenAI signals 'memory > prompt' direction - Google Gemini 3 Deep Think adds multi-hypothesis 'multi-threaded' reasoning

Key insights

Industry report: agentic reasoning becomes mainstream

The a16z and OpenRouter 'State of AI 2025' report announced a structural shift from single-turn generation to multi-step reasoning and agentic behavior. The report used OpenRouter’s 100 trillion-token interaction dataset to show reasoning-capable models (citing OpenAI o1) now perform planning, tool use, and multi-step problem solving. The report stated reasoning models account for over 50% of model traffic and that average prompt lengths have increased roughly fourfold, indicating users supply far more context and expect multi-step outputs.

User behavior: longer prompts and routine tool calls

OpenRouter data in the report showed users now provide much larger context to models and accept longer response latencies for deeper reasoning. The report documented a roughly 4x increase in prompt length and widespread model tool invocation (e.g., web search, code execution), describing tool-calling as a normalized workflow for high-value tasks. It also highlighted that users prefer models that can plan and use external tools rather than only producing instant single-turn answers.

Open-source landscape: mid-size models and Chinese entrants gain share

The report described an expanded open-source ecosystem where mid-size models (15B–70B parameters) became the preferred tradeoff between cost and capability. It named Chinese open models—DeepSeek, Qwen, Moonshot, ZAI—as rising competitors that grew from marginal presence in late 2024 to about 30% of the open-source market by 2025. The report framed the current market as multi-polar, with multiple organizations competing rather than a single dominant open model.

Use-case concentration: programming and roleplay dominate

OpenRouter usage statistics reported programming as the fastest-growing and most strategically significant application, with heavy embedding of models into developer IDEs for code writing and bug fixing. The report also identified roleplay as the single largest open-source application, making up roughly half of open-source usage due to its uncensored, customizable affordances for creative writing and immersive interaction.

Architectural advance: gating improves attention and long-context performance

A NeurIPS 2025 best-paper reported systematic experiments on gating mechanisms added to Transformer attention. The study evaluated over 30 gating variants up to 15B parameters and found that inserting learned gates per head before concatenation yielded the best results. Reported benefits included more stable training enabling higher learning rates, reduction of early-token 'attention sinks,' and significant improvements on long-text benchmarks when context windows were expanded.

Vendor product directions: memory and multi-threaded reasoning

Vendor signals align with the reasoning trend: OpenAI’s recent blog emphasized shifting emphasis from prompt engineering toward persistent memory, stating models will increasingly learn user habits. Google’s Gemini 3 introduced a 'Deep Think' mode that explores multiple hypotheses concurrently, described as multi-threaded reasoning. These vendor updates indicate major providers are prioritizing memory and structured multi-step reasoning features.

Overall implication: competition now on complex problem solving

The combined findings and vendor updates show the competitive axis shifting from surface-level conversational fluency toward models' ability to perform long-chain reasoning, plan, use tools, and handle long context. The report and accompanying research note that innovation will focus on reasoning architectures, memory systems, and multi-step agentic workflows.

Source posts