State of AI Report 2025 is an annual open report produced by Nathan Benaich and the Air Street Capital team. It spans research, industry, politics, safety, a practitioner survey, and predictions. It is effective at organizing scattered events into an industry narrative, but it is not a neutral statistical yearbook or a peer-reviewed paper.
The best way to use it is as a map of hypotheses worth checking: first understand which developments the authors connect, then return to primary papers, company disclosures, and public data for individual claims.
Four themes
The report places models, products, infrastructure, and policy in one industry chain. The themes interact, but they use different kinds of evidence and should not be read with equal confidence.
| Theme | Report's assessment | Reading limit |
|---|---|---|
| Reasoning models | Test-time compute and post-training became central to frontier competition | Benchmark contamination, variance, and prompt differences can magnify apparent progress |
| Open weights | China's open-weight ecosystem accelerated while leading closed models retained capability-per-cost advantages | Open weights do not imply open training data or full reproducibility |
| Agents | Coding, science, and computer-use tools began producing practical value | Long-horizon reliability, transfer, and cost still constrain production use |
| Infrastructure | Multi-gigawatt clusters entered planning and electricity became a roadmap constraint | Plans, built capacity, and actual utilization must be separated |
Reasoning became a product capability—and a measurement problem
The report characterizes 2025 model competition as a race around think-then-answer systems, with OpenAI, Google, Anthropic, and DeepSeek trading leads. More test-time compute and reinforcement-learning post-training moved reasoning directly into products.
It also presents challenges to claimed reasoning gains: results can sit within baseline variance, small changes to problems can cause sharp degradation, and a chain of thought may not faithfully represent internal processing. A capability demonstration and a repeatable measurement are not the same thing.
Open-weight competition accelerated, but openness is not one property
Chinese models including DeepSeek, Qwen, and Kimi helped accelerate the open-weight ecosystem and lowered the barrier to using strong models. The report nevertheless assesses the strongest closed models as retaining a capability-per-cost advantage.
Comparisons need to separate downloadable weights, code, training data, licensing, and reproducibility. Calling every open-weight release fully open source hides meaningful differences in deployment freedom, safety review, and supplier dependence.
Agents became useful, while reliability still sets the boundary
Tools for coding, research, medicine, and computer use moved from demonstrations into concrete workflows. Tool calls, environment feedback, and task decomposition allow models to complete longer sequences of work.
The report also tracks transfer failures, reward hacking, long-context limits, memory problems, and cost. Completing one demonstration does not establish reliable operation under controlled permissions, changing inputs, and recoverable failures.
Electricity became a roadmap constraint
Multi-gigawatt data-center plans, grid connection times, and power geography began shaping training and inference deployment. Compute competition now includes electricity, construction schedules, cooling, and capital structure—not only chip procurement.
Planned capacity is not operating capacity. Infrastructure claims need to distinguish announcements, sites, construction, grid connection, and actual utilization, and to separate the report's industry interpretation from project-level facts.
How to use this report
The report is useful for finding cross-domain connections and questions for further investigation, not as a single source of factual authority. Its venture-investor authorship provides industry access and perspective while also requiring attention to investment framing and selection bias.
For product decisions, convert its conclusions into verifiable questions: Can the capability be reproduced? Does cost include the full reasoning budget? Is the infrastructure operating? Has the agent passed a real workflow with controlled failure modes?
Methodology
This original synthesis draws on the report's executive summary and six thematic sections and cross-checks the edition against the official State of AI 2025 page.
State of AI is an analytical industry report rather than a statistical yearbook with one methodology; assessments are attributed to the report and are not presented as independent AI3Radar measurements.
Lead author Nathan Benaich is a general partner at Air Street Capital; this relevant interest is disclosed as a reading constraint.
This article reproduces no report figures, treats no prediction as fact, and does not claim to have replicated experiments or project capacity.