2025 in AI: twelve months that changed how software gets made, how we search, and how countries compete
It is a cliche to say that a year in technology moved faster than any before it. It is also true that some years genuinely represent a step change rather than incremental progress. 2025 was one of those years for AI. Looking back across twelve months, the scale of what changed is difficult to fully appreciate in real time. This is an attempt to catalogue the moments that actually mattered and say something honest about what they add up to.
The year at a glance
| Item | Value |
|---|---|
January DeepSeek-R1 released Challenged the cost assumptions of frontier AI | |
February Vibe coding coined, Paris AI Summit Named a new paradigm; exposed governance fractures | |
March Claude 3.7 extended thinking Reasoning models became mainstream | |
April Meta Llama 4 released Open-weights models reached GPT-4o parity | |
May Google I/O 2025, Claude 4 7-hour session AI Mode reimagined search; agents went autonomous | |
June Humanoid robots in Amazon fulfilment Physical AI crossed from demo to deployment | |
July AI content flood recognised Web quality and model training threatened | |
August GPT-5 launch, Grok Imagine expansion PhD-level benchmarks met; content policy tested | |
September Coding agents production audit Honest review of agent wins and failure modes | |
October Multimodal AI matures Voice, video, image understanding all production-ready | |
November Mission Genesis signed AI formalised as national security priority | |
December Year-end reflection Pattern recognition across a transformative year |
The economic story: DeepSeek to Mission Genesis
The year started with DeepSeek-R1 demonstrating that frontier-level reasoning might be achievable at a fraction of the expected cost. It ended with Mission Genesis committing more money to AI infrastructure than any previous peacetime government technology programme. These two facts are not in contradiction. If anything, DeepSeek made Mission Genesis more urgent from a policy perspective: if efficient training methods spread globally, US spending advantages matter less, and competitiveness depends on other factors like data, talent, and application deployment speed.
The practical economic effect was that API costs fell dramatically across the year. Tasks that cost one dollar per thousand queries at the start of 2025 cost a tenth of that or less by December. The economics of building AI-powered applications changed from expensive experiments to routine product decisions.
The developer story: vibe coding to autonomous agents
February's vibe coding moment and September's coding agent production audit bracket a year in which how software gets built changed more than in any previous year. The combination of capable language models, IDE integration, and agentic capabilities meant that a single developer could accomplish what previously required a small team.
This was not straightforwardly good news for everyone. Junior developer roles, particularly those focused on boilerplate and implementation of clearly specified features, faced compression. The value of developers who could specify tasks clearly, review AI output critically, and understand system-level constraints increased. The market for undifferentiated coding work declined. These transitions are not catastrophic, but they are real and fast.
The physical world story: humanoids arrive
The deployment of humanoid robots in Amazon fulfilment centres in June was not a flashy moment. There was no product launch event, no keynote. But it represented something the robotics industry had been promising for years: AI-capable physical systems handling real tasks in real production environments at a scale that could affect hiring decisions.
The LLM-enabled perception systems that make these robots useful are the same systems being developed for autonomous vehicles, surgical assistance, and industrial inspection. 2025 was the year the path from text-based AI to physical AI became visibly clear and much shorter than expected.
The governance story: a year of fractures
The Paris AI Summit in February, Mission Genesis in November, and the scattered national AI strategies in between told a consistent story: the world is not going to agree on how to govern AI. The US and UK refused to sign the Paris declaration. The US then framed AI explicitly as a national security competition. The EU continued with its AI Act implementation, which most large US AI companies were working to comply with while lobbying to soften specific provisions.
There was no unified international governance of AI in 2025 and there will not be in 2026. The realistic alternative is a set of overlapping, partially compatible frameworks that create compliance costs for global companies and inconsistent experiences for users depending on their jurisdiction. That is messy, but it is how most technology governance has worked historically.
What 2025 added up to
- 10x
- API cost compression across the year
- 3+
- Reasoning models reaching human expert level
- 7 hrs
- Longest autonomous AI coding session
- $500B+
- US AI infrastructure committed
The honest summary is that 2025 was the year AI moved from being an impressive technology to being infrastructure. The tools were reliable enough, cheap enough, and capable enough to build businesses on. The question was no longer whether AI could do useful things. The question was which specific applications were worth building, who should bear the costs and risks of deployment, and how to handle the consequences for work, creativity, and governance that were now visibly arriving rather than hypothetically approaching.
The technology did not slow down in December. The year ended with capabilities that would have been considered frontier-only at the start of the year now available for free or close to it, with more powerful systems already in training at the labs. 2026 picked up where 2025 left off.