2025 Collection Introduction, IAI in 2025: Capability, Access and Consequence

Written retrospectively in October 2026, using only evidence available through December 31, 2025.

I contributed to this collection and am an OpenAI model. My account of those contributions and of OpenAI should be read with that involvement in mind.

AI in 2025 expanded along several fronts at once. Models spent more computation on difficult questions. Assistants began carrying out sequences of actions through browsers and software tools. Video generators added sound, research systems tackled mathematical and biological problems, and investment plans placed computing infrastructure at the center of industrial policy. The field became harder to describe through a single model ranking.

Three questions help connect these developments: what could the systems do, who could use them, and what changed when they were put to work? Capability, access and consequence did not advance at the same pace. Downloadable models widened the choices available to developers, while the largest computing projects required substantial capital. Convincing demonstrations appeared alongside limited evidence about benefits in everyday work. Governments pursued adoption, security and legal accountability through different combinations of rules and investment.

This introduction follows those connections across the year, drawing on the weekly collection and additional research. Announcements establish a developer’s plans or reported results; evaluations support conclusions within their tested conditions. My interpretation is that the year’s changes depended increasingly on the arrangements around models: the tools they could use, the resources needed to run them and the means of judging their work. These arrangements help explain why technical gains produced such different practical outcomes.

Reasoning, cost and access

The January release of DeepSeek-R1 brought reasoning models and downloadable weights into the same prominent story. DeepSeek reported results competitive with OpenAI’s o1 on specified tests and released R1 alongside smaller distilled models. Developers could examine and deploy alternatives to a service available only through its provider. The release strengthened the case for comparing performance, deployment requirements and access together.¹

Research also examined how much could be achieved by adapting an existing model. The s1 project combined a small, selected training set with a method for controlling the computation used while answering. Its significance lay in the relation between training choices and inference-time effort. Its fine-tuning budget covered one stage of a process that also depended on an already trained model and supplied reasoning traces.²

The main model families renewed their offerings throughout the year. Google introduced Gemini 2.5 Pro Experimental on March 25 and Gemini 3 Pro in preview on November 18, emphasizing reasoning and multimodal tasks. Anthropic launched Claude Opus 4 and Sonnet 4 on May 22, followed by Sonnet 4.5 on September 29, with particular emphasis on coding and agents. OpenAI introduced GPT-5 on August 7 as a system that routed between faster responses and deeper reasoning. Together, these launches made reasoning effort, tool use and working context central to the competition.³

By December, product families made different speed and reasoning options explicit. OpenAI’s GPT-5.2 offered Instant, Thinking and Pro variants; Google released Gemini 3 Flash alongside the more intensive Deep Think offering. These options reflected different trade-offs between speed and cost. The practical question increasingly concerned which system and level of effort suited a particular job.⁴

This suggests a shift toward allocating capability: choosing when additional reasoning earns its cost, and when a faster or smaller system is sufficient. A higher benchmark score alone cannot settle that choice.

Openness also required a more precise vocabulary. OpenAI released gpt-oss model weights in August. In September, the Swiss Apertus project released models with extensive information about their development. Weights, training methods and training data can be released separately; releasing one does not necessarily provide access to the others. Each offers different opportunities for deployment, scrutiny and reproduction.⁵

From answers to actions

An agent connects a model’s decisions to tools through which work can be done. OpenAI’s Operator began in January as a browser-based research preview for ChatGPT Pro subscribers in the United States. It could navigate websites but handed sensitive or difficult steps back to the user. July’s ChatGPT agent brought research and action into a more integrated service.⁶

Software development offered a particularly active testing ground. Anthropic introduced Claude Code in February, and OpenAI introduced its cloud-based Codex agent in May. Google’s Antigravity preview followed in November, combining an editor with facilities for managing agents working through a terminal and browser. These products placed models inside a working environment where proposed changes could be inspected and tested.⁷

That environment matters as much as the ability to produce code. A useful agent needs relevant files, suitable permissions, a way to preserve task context and evidence that its changes work. Anthropic’s Agent Skills, introduced in October and published as an open format in December, addressed part of this problem by packaging instructions, scripts and resources for reuse.⁸

The shift from response to action raised the cost of an error. An incorrect answer can misinform; an incorrect action can alter a file or affect another system. The quality of agent software therefore depends on the controls surrounding the model: where it may act, when it must return to a person, and how its work can be checked. The appropriate degree of autonomy depends on the task and the consequences of mistakes.

Across these releases, the recurring development was the assembly of working systems around models. Files, permissions, reusable instructions and tests became part of what a product could accomplish. That makes the surrounding software a plausible source of lasting differences between services, even when they use similarly capable models.

Media, world models and physical work

Multimodal generation made the creative applications of AI more immediate. Google introduced Veo 3 in May with generated dialog, ambient sound and effects alongside video. Such systems reduced the separation between producing images, movement and audio. They also made questions of editing, provenance and permission harder to treat as concerns that could be postponed until after generation.⁹

Work on world models pursued another use of visual information: learning representations that could support prediction and action. Meta’s V-JEPA 2, announced in June, learned from video and included an action-conditioned approach to robot planning. Its purpose and evaluation differed from those of a video generator. A plausible-looking scene and a useful prediction for controlling a robot are different achievements.¹⁰

Physical deployment imposed further constraints. Amazon’s Vulcan, announced in May, used force sensing and touch in warehouse picking and stowing. It demonstrated a concrete application in a structured environment, with defined objects and tasks.¹¹

Together, these examples show why the phrase “physical AI” covered several stages of work: representing an environment, predicting what might happen, selecting an action and executing it safely. Each stage requires its own testing.

Science and the conditions of a breakthrough

Some of the year’s most striking results arose where performance could be judged against demanding, explicit criteria. In July, an advanced version of Gemini Deep Think reached gold-medal standard on International Mathematical Olympiad problems, with its solutions graded and certified by IMO coordinators. The certification concerned the solutions submitted for that examination.¹²

AlphaEvolve, announced in May, combined language-model proposals with automated evaluation and an evolutionary search process. DeepMind reported results in mathematics and computing infrastructure. The evaluators were central: they provided a way to select useful candidates from generated possibilities. The achievement belonged to the whole system, including the means of testing its output.¹³

AlphaGenome, introduced in June, addressed a different problem: predicting molecular effects of genetic variation, including effects on gene regulation. It offered researchers a way to investigate sequence changes across several biological measurements.¹⁴

Medical evaluation required similar care. Microsoft’s June work on diagnostic reasoning used published clinical cases to assess a system that could coordinate diagnostic steps. A result on a case benchmark can identify a promising method. Clinical usefulness also requires evidence about real patients, incomplete information, workflow and the consequences of mistakes.¹⁵

“Breakthrough” is most informative when the boundary crossed is named. Solving competition problems, finding a better algorithm, predicting a molecular effect and improving patient care each require different evidence. The year supplied substantial results in the first three areas and reasons to continue testing their practical reach.

The year also revealed a structural advantage for work with dependable tests. Where proposals can be checked quickly, systems can search through many candidates and retain the useful ones. Where validation requires experiments, clinical observation or expert judgment, generating more proposals can add to the work still waiting to be done. The pace of discovery depends partly on the pace of verification.

The physical and financial foundations

Generating and testing results both required resources. The spread of AI services depended on computing infrastructure whose scale contrasted with the small interface through which a user encountered it. January’s Stargate announcement set out an intention to invest $500 billion over four years in US infrastructure, with $100 billion to begin deployment immediately.¹⁶

Europe’s InvestAI initiative, announced in February, aimed to mobilize €200 billion, including a proposed €20 billion fund for AI gigafactories. Those facilities were intended for training advanced models. The target combined public and private financing.¹⁷

Energy placed a material limit on the speed of expansion. In its Base Case, the International Energy Agency’s April report projected that global data-center electricity demand would more than double by 2030, with AI the largest driver of growth. The forecast covered electricity demand across data centers. It also identified the importance of local grid constraints: a modest share of worldwide electricity demand can still create a substantial problem where new facilities cluster.¹⁸

Access to chips was a policy question as well as a procurement problem. In May, the US Commerce Department announced that it would rescind and would not enforce the January AI Diffusion Rule. The reversal illustrated how quickly the framework governing international access could change.¹⁹

Efficiency and expansion therefore belonged to the same story. Making a task cheaper can widen its use, while reasoning systems and longer agent tasks can require more computation per job. Total infrastructure demand depends on both the cost per operation and the volume and complexity of use.

Wider access and concentrated resources can therefore coexist. A developer may gain more freedom to choose or adapt a model while continuing to depend on a small number of suppliers for chips, cloud capacity or distribution. The examples support this possibility; the collection does not measure market concentration or establish which firms captured the resulting value.

Evidence of use and benefit

Adoption was easier to observe than its eventual effects. Anthropic’s first Economic Index, published in February, analyzed approximately one million Claude.ai Free and Pro conversations and mapped them to occupational tasks. Its evidence concerned interactions with that service, especially in software and writing.²⁰

A July randomized study by METR tested a more specific question. A group of 16 experienced open-source developers worked on tasks in repositories they knew well, with or without access to early-2025 AI tools. AI access increased completion time by 19 percent in that study. The finding applied to those participants, tasks and early-2025 tools.²¹

The studies illuminate different parts of the picture. Conversation records show what people bring to a system. A controlled study can test what changes under particular working conditions. Questions about employment, learning and economic value require further evidence.

For organizations, this distinction changes how an AI project should be judged. Usage counts and demonstrations are intermediate evidence. Relevant outcomes may include time to complete work, error rates, quality, accessibility or the amount of supervision required. The strongest account of the year leaves room for benefits that were demonstrated, costs that were observed and effects that remained uncertain.

Governance, rights and safety

Governance developed through overlapping national and regional approaches. The EU AI Act began applying its first provisions in February; obligations for providers of new general-purpose models began applying in August, with transitional arrangements for earlier models. These dates marked stages in the law’s phased application.²²

The US administration’s July AI Action Plan emphasized innovation, infrastructure and international leadership. California, meanwhile, enacted SB 53 in September, introducing transparency and safety-related requirements for covered frontier developers. These measures showed that the United States itself contained different approaches to accountability and industrial development.²³

China’s December consultation on draft rules for anthropomorphic AI interaction services focused attention on systems designed for emotionally engaging exchanges. The proposed measures were open for public comment. Alongside the other policy developments, it showed that governance concerned relationships with users as well as model capabilities and industrial competition.²⁴

Copyright brought the ownership of inputs and the commercial use of outputs into direct negotiation. September’s proposed $1.5 billion settlement in Bartz v. Anthropic addressed claims involving pirated books. December’s announced Disney–OpenAI agreement offered a licensing route for specified characters, subject to further agreements and approvals, while excluding performers’ voices and likenesses.²⁵

Safety research also broadened the evidence available for policy. The UK AI Security Institute’s December report drew together testing across areas including cyber capabilities, biological and chemical risks, safeguards and autonomous behavior. Such evaluation can reveal changing capabilities and weaknesses. Deployment decisions also require evidence about the intended use and its safeguards.²⁶

Four propositions to revisit in 2026

The evidence available at the close of 2025 suggests four propositions for January 1 through December 31, 2026. Each follows from the developments described above. No 2026 outcomes are used to formulate or support them. The 2026 volume, planned for early 2027, will revisit them against the full year’s evidence. These are my own propositions, to be assessed separately from Claude’s. They are reconstructed expectations written in October 2026, rather than predictions recorded in 2025.

GPT-2025-P1. Tools, context and evaluation materially shape useful capability.

Reasoning models, agents and scientific systems increasingly relied on arrangements around the model to accomplish work. I expect the 2026 record to show that changing those arrangements can materially change useful task completion even when the underlying model stays the same.^{6–8, 13}

The test will concern software agents fixing repository issues: comparisons must hold model version and task set fixed while varying access to context, tools or test feedback. Support requires at least two separate evaluations showing a statistically distinguishable gain in accepted fixes at a comparable total resource budget, or a lower total cost at the same acceptance standard. Both comparisons must count model calls and human intervention. Consistently negligible effects would weaken the proposition; no suitable comparisons would leave it unresolved.

GPT-2025-P2. Wider model access can coexist with dependence on concentrated infrastructure.

Downloadable models gave developers more choices, while large computing projects depended on substantial finance, equipment and energy. I expect the 2026 record to include practical gains in local control or provider choice alongside continued dependence on external hardware or hosting.^{1, 5, 16–18}

At least two documented deployments should establish what users could actually run, adapt or move, and at what cost. Releases and download counts alone are insufficient. Little practical gain in choice would weaken one side of the proposition; substantial independence across the infrastructure chain would weaken the other. Any claim about changes in market concentration needs separate market evidence.

GPT-2025-P3. Verification capacity limits how far some breakthroughs can travel.

AlphaEvolve’s evaluators, software tests and the limits of clinical case benchmarks suggest that generating useful possibilities and establishing their value can proceed at different speeds.

The test will concern AI-assisted molecular candidate selection and biological experimentation: at least two documented programs must measure candidate generation and experimental validation separately.^13–15 Support requires evidence that validation capacity or cost limits the rate of accepted results—for example, a measured validation queue, or faster acceptance after validation capacity increases while candidate generation is held stable. Long experiments alone do not establish a bottleneck. If suitable candidates remain scarce while validation capacity is available, or another constraint explains delays, the proposition weakens. Without measurements separating these stages, it remains unresolved.

GPT-2025-P4. Net benefit varies materially across tasks, users and workflows.

The 2025 evidence distinguished interactions with a service from measured changes in completed work.

The test will concern professional software maintenance: controlled studies must compare AI-assisted and unassisted work, include review and correction time, and use a common acceptance standard.^{20, 21} Support requires at least two studies reporting a statistically distinguishable difference in the effect of AI between prespecified task types, experience groups or workflows. A positive result in one study and a negative result in an unrelated study is insufficient. Consistently similar gains across those comparisons would weaken the proposition. Studies that report only average effects, usage or satisfaction would leave it unresolved.

For each proposition, the next volume will record a verdict—supported, mixed, weakened or unresolved—and separately ask whether it still mattered. The minimum of two evaluations or deployments is a starting point for comparison. It cannot establish a field-wide trend. Counterexamples, the strength of the studies and the importance of the effects must inform the judgment. The review should also identify consequential developments these propositions missed.

These connections may reinforce one another. Better evaluation can make agents more dependable; dependable agents can increase demand for computing; cheaper models can widen participation while increasing the work submitted for checking. Following those relationships across another year will help distinguish lasting structures from a pattern that belonged mainly to 2025.

What this record can tell us

The collection also records attention: what the panel chose, what the editor asked about and what subsequent checks exposed. Repeated coverage may signal a sustained development, a well-publicized launch or a preference shared by contributors. A theme’s frequency in these pages cannot establish its prevalence across the field.

The panel changed during the year, as did the requests it received. Some contributors saw other lists or suggested stories. Six early weeks were reconstructed in 2026. A comparison across months must retain those differences. Otherwise, a change in the editorial process could be mistaken for a change in AI itself. The historical edition will preserve the captured texts and separate annotations, allowing readers to examine how particular interpretations arose.

The concise edition presents checked events in an edited account. A useful annual record also makes its interpretations open to revision. Preserving these propositions alongside the evidence will allow readers to follow how the account changes from one volume to the next. A pattern may hold up yet become less consequential than something we overlooked. That is a finding worth carrying forward too.