2026年人工智能状况：半年快照

January 2026 State of AI: Bi-Annual Snapshot The Execution Era of AI Introduction We believe that building and operationalizing AI products is no longer just thenew frontier of competitive advantagebut ratherbecoming table stakes in the software world.In Q2 2025, we published “The AI Builder’s Playbook” to elevate the voices of thearchitects, engineers, and product leaders driving this work and emphasize what it takes to conceive, deliver, and scale AI-powered Six months later, the picture is clearer.Over the last six months, we believe the AI market has entered a new phase of maturity.Whatstarted as the race to experiment with large models and launch early AI features has increasingly evolved into a challenge ofscaling AIinto durable, economically sound products.Given the speed of evolution in this market, this report is designed as a bi-annual update This report revisits core dimensions of the builder’s playbook, highlighting the most importantchangesand developmentsover thelast six months. Grounded in our proprietary Q2 2025 and Q4 2025 surveys of executives at software companies building AIproducts, alongside perspectives from our ICONIQ Community, the 2026 State of AI report seeks to offer a longitudinal operatorperspective on what it takes to turn AI from a capability into a durable competitive advantage. In our view, the findings pointto a Explore Our AI Perspectives DataSources & Methodology This study summarizes data from aQ4 2025survey of ~300 executives2at softwarecompanies building AI products, including CEOs, Heads of Engineering, Heads of AI,Headsof Product, Chief Revenue Officers, Throughout this report, wecompareinsights to our prior State of AI report,published in Q2 20253, “The AI Builder’sPlaybook”, where applicable. Wherenecessary, longitudinal data has beennormalized to account for differences infirmographicstoensure trends are We also weave ininsightsand what webelieve to be best practices from AI leaders All industry perspectives shared in thisreport have been anonymized to protectcompany-level information. From Models to Products: Where We See AI Differentiation Being Built We believe that AI product development has entered a phase of standardizationand maturity. As the base models continue to improve, builders are no longerfocused on creating foundational models but instead ondelivering differentiatedproducts at the application layer. Nearly70% of companies are building vertical AIapplications, reinforcing that durable value is being created through domain-specific workflows rather than generalized intelligence. Consistent with this shift, As model quality continues to improve across providers, our survey showsbuilders are increasinglyadopting multi-model strategies to balance reliability, cost, latency, and customization. On average, companiesnow leverage ~3.1 modelproviders, up from ~2.8 six months ago, reflecting a growing emphasis onorchestration rather than allegiance to a single platform. However, despiteincreased investment in data pipelines and evaluation, most companies still reportthat their data foundations are only “mostly” or “somewhat” ready, particularly at Application layer products continue to be the most common types of products being developed by AI builders, with almost As base models evolve and improve in efficacy, it appears application layer innovation is the primary differentiator for AIbuilders, competing on product UX, workflows, and integrations rather than proprietary model development From Models to Products: Where We See AI Differentiation Being Built Application-focused builders most heavily rely on third-party model APIs, while proprietary model developers tend to Model Providers by Primary Differentiator% of Respondents, Select All That Apply, N=202 From Models to Products: Where We See AI Differentiation BeingBuilt Top model selection criteria have remained consistent over the last 6 months, pushing builders toward multi-model Builders are Focusing on Model StackEfficiency At ICONIQ’s recent forum for enterpriseChief Data and AI Officers, leadersdiscussed their increasing focusonshifting to a cost-efficient model stack.Leaders have emphasized that frontiermodels are often unnecessary for mostautomation tasks and that open-source Additionally,routing strategies areemerging:the majority oftasks arepushed to smaller models, with only OpenAI remains the most widely used model provider among survey respondents; however, builders are using a widervariety of models over time. Notably, Gemini has increased to the second most popular provider since our Q2 2025 survey To measure performance of AI models, builders are adopting multiple evaluation methods; however, evaluation remainslargely user feedback-driven and manual today, with only 52% of builders adopting automated eval frameworks From Models to Products: Where We See AI Differentiation BeingBuilt Earlier-stage products tend to rely on manual controls to reduce hallucinat