Free tools Windows power users keep installed
One-click scans. No signup required.
The most accurate answer to “what’s next for AI in 2025?” is a set of trends, not a single forecast. In 2025, AI systems kept improving on demanding tests, the cost of running a given level of capability fell sharply, companies produced most of the notable new models, governments added rules at a faster pace, and institutions were not evenly ready to put the technology to use. Written in October 2026, this article looks back at what the 2025 evidence showed, separates measured results from predictions, and marks where the evidence stops.
How to read the 2025 evidence
Most of the figures below come from Stanford HAI’s annual AI Index and from OECD government reports. They do not cover the same periods, so check the year behind every number before you repeat it. The 2025 AI Index was published in April 2025 and mostly describes 2024. The 2026 AI Index looks back at technical performance during 2025.
| Source | Published | Period covered | Used here for |
|---|---|---|---|
| Stanford HAI, “AI Index 2025: State of AI in 10 Charts” | April 7, 2025 | Summary of the 2025 AI Index; several figures describe 2024 | Overview of the main trends |
| Stanford HAI, Artificial Intelligence Index Report 2025 | 2025 | Mainly 2024 | Inference costs, industry share, regulation counts, education |
| OECD, Governing with Artificial Intelligence: The State of Play and Way Forward in Core Government Functions | June 2025 | Dozens of governance approaches and 200 government AI use cases | Public-sector maturity and examples |
| OECD, “Enablers, guardrails and engagement for unlocking trustworthy AI” | 2025 | Not stated | Incremental use cases and limited dedicated government AI laws |
| Stanford HAI, The 2026 AI Index Report | 2026 | 2025 | Technical performance and broader chapters on economy, governance, education, science and medicine |
None of these reports is a forecast of 2026 or later. They describe what was measured, so the trends below are best read as direction rather than destiny.
Capability kept improving, as a trend
The 2025 AI Index reported that performance on demanding benchmarks continued to improve. The 2026 AI Index describes technical performance during 2025 across language, image, video, speech, reasoning, robotics and agentic systems, meaning systems that carry out multi-step tasks on a user’s behalf.
#1 Best Overall
A benchmark is a controlled test. A higher score shows progress on that test. It does not show that a tool will be reliable on an open-ended task, such as reading a loan agreement or sorting a month of bank transactions. Those tasks depend on the specific input, the checks around the output, and the person reviewing it.
AI got much cheaper to run
The clearest cost signal comes from the 2025 AI Index. It reports that the inference cost of a system performing at GPT-3.5 level fell by more than 280-fold between November 2022 and October 2024. Inference is the cost of running a trained model to answer a request, as opposed to the cost of training it. The report attributes much of the decline to increasingly capable small models.
Rank #2
What the figure measures
- It tracks a fixed capability level, GPT-3.5 level, rather than the newest model on the market.
- It is a cost-of-running measure. It is not the subscription price or per-use fee a consumer sees.
- It covers a window that ends in October 2024.
What it does not mean for your bill
A lower cost per unit of capability does not automatically mean lower prices for end users. A provider can pass savings on, keep them as margin, or spend them on larger models and more usage. The sources do not establish how far the savings reached consumers, so treat cheaper inference as a trend on the provider side, not a promise about what you will pay.
A budget check for AI tools
- Confirm the billing basis: monthly fee, usage-based charge, or a bundle with another service.
- Check what a free tier limits, such as features, daily usage or file size, and when those limits reset.
- Read the current pricing page on the day you sign up. Prices for AI tools change, and a figure from a 2025 article may be out of date.
- Note renewal dates and cancel or downgrade any subscription you no longer use.
Industry built most notable models
Nearly 90% of notable AI models in 2024 came from industry, up from 60% in 2023, according to the 2025 AI Index. Academia remained the leading source of highly cited research in the same report. These figures measure different things: who produced the models Stanford HAI counted as notable, and whose research other researchers cite most. Neither measure forecasts who will lead in later years.
Governance grew, unevenly
U.S. federal activity
In 2024, U.S. federal agencies introduced 59 AI-related regulations, more than twice the number introduced in 2023, according to the 2025 AI Index. That count shows how much regulatory activity there was. It does not show how many of those regulations took effect, how they interact, or whether they work as intended.
Legislative activity across 75 countries
Legislative mentions of AI across 75 countries rose 21.3% since 2023, according to the 2025 AI Index. A mention count tracks how often AI appears in legislative activity. It is not a count of enacted laws, so it cannot tell you how many countries have binding AI rules.
Government use
The OECD’s June 2025 report analyzed dozens of governance approaches and 200 AI use cases in core government functions. Most of the applications it analyzed were incremental improvements and productivity gains, according to the OECD, and dedicated AI laws or regulations for government use remained limited. These findings describe government use the OECD studied. They do not describe every public or private sector.
Independent reader supportYour contribution helps us test, update, and keep practical guides available for everyone.Readiness lagged access
Education
Two-thirds of countries offered or planned K–12 computer-science education, according to the 2025 AI Index. Offering a course is not the same as having trained teachers or equal access. The report describes access and teacher readiness as uneven.
Best Value
Government maturity
The OECD’s June 2025 report found practical examples from which governments can learn. It also found that AI maturity was not yet prevalent in governments, and it pointed to gaps in the enabling conditions that make AI projects workable, such as skills, infrastructure and capacity to deploy systems responsibly.
The trade-offs in one view
| Axis | Why it matters to readers | What the evidence does not show |
|---|---|---|
| Capability versus cost | Better performance and lower running costs are separate developments. A cheaper tool is not automatically a better one for your task. | Whether savings reach end users in lower prices |
| Industry versus academia | The company that releases a model is not necessarily the group that produced the research behind it. | Which sector will lead in later years |
| Innovation versus governance | More rules show more attention to AI, not more enforcement. | Whether rules are effective or applied consistently across jurisdictions |
| Access versus readiness | A model you can use is not the same as a school, workforce or agency that can use it well. | Whether benefits spread evenly or quickly |
What to watch in future editions
These are the measurements the sources track, so they are useful for checking whether the 2025 trends continue. None of them is a prediction.
Quick Recap
- Cost per unit of capability in future Stanford AI Index editions, and whether it is reported next to consumer prices.
- Counts of enacted AI rules, not only introduced ones or legislative mentions.
- Follow-up OECD work on how many governments have dedicated AI laws for their own use.
- Education data on teacher readiness, not only on whether a country offers computer-science courses.
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
