10 Major AI Innovations of 2025

Looking back at a year is easy to do badly. Memory favours whatever got the loudest launch, and launch posts are written to impress. This list takes a narrower route: ten developments from 2025 that we could confirm on a primary page, each with a date, a plain description of what it was and one limit to keep in mind.

What are the ten, in date order?

1. DeepSeek-R1: reasoning from reinforcement learning, with open weights

On January 22, 2025, DeepSeek-AI posted the paper DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning to arXiv. Its abstract argues that reasoning abilities can be incentivised through reinforcement learning alone, without human-labelled reasoning examples, and reports emergent behaviours such as self-reflection and verification. The model card for the released model says the code and weights are licensed under the MIT License, and lists a 671-billion-parameter mixture-of-experts model (37 billion active) plus six smaller distilled models built on Qwen and Llama bases.

Why it matters: it put a reasoning-focused model in public hands under a permissive licence, which let others study and build on it. Limit: the paper's abstract itself does not describe the release, so the open-weight claim rests on the model card. 'Open weights' also does not mean open training data or code. See open-weight models explained.

2. Agent2Agent: a protocol for agents to talk to each other

On April 9, 2025, Google announced the Agent2Agent (A2A) protocol, described in its post as 'an open protocol that provides a standard way for agents to collaborate with each other'. The post names more than 50 technology partners at launch, including Atlassian, Box, Salesforce, SAP and ServiceNow, and presents A2A as complementary to Anthropic's Model Context Protocol.

Why it matters: if agents from different vendors are to hand work to one another, they need a shared way to describe tasks and results. Limit: a published protocol is a proposal until people adopt it. The partner list shows intent, not usage.

3. Ironwood: a TPU positioned for inference

Also on April 9, 2025, at Google Cloud Next 25, Google announced Ironwood, its seventh-generation Tensor Processing Unit. Google describes it as the first TPU designed specifically for inference, meaning running trained models, and says it will be offered in 256-chip and 9,216-chip configurations. The post lists many performance and efficiency figures; those are Google's own claims and we do not repeat them here as established.

Why it matters: it reflects a wider shift of spending from training models to serving them, particularly the longer outputs of reasoning models. Limit: all performance statements come from the vendor, and we did not test the hardware. For the wider chip picture, see our article on AI hardware and chip companies.

4. Claude Code reaches general availability

On May 22, 2025, Anthropic announced Claude Opus 4 and Claude Sonnet 4 and said Claude Code, its command-line coding agent, was generally available. The post describes background tasks through GitHub Actions, beta extensions for VS Code and JetBrains, and a beta GitHub integration that lets users tag Claude Code on pull requests to respond to review feedback or fix CI errors.

Why it matters: it was one of the clearer examples of an assistant that works across a whole codebase and takes actions, not just suggests text. Limit: an agent that edits files and runs commands needs review and permission controls. Read what an AI coding agent is before giving any agent access to a repository, and note that this site publishes no test of Claude Code.

5. The EU's rules for general-purpose AI models begin to apply

The European Commission's AI Act page says the Act entered into force on August 1, 2024, and that its governance rules and obligations for general-purpose AI (GPAI) models became applicable on August 2, 2025. It also records that the Commission published a GPAI Code of Practice in July 2025, which it describes as a voluntary compliance tool prepared by independent experts, covering transparency, copyright, and safety and security.

Why it matters: it was the first time a major jurisdiction's binding duties for the underlying models, not only for applications, took effect. Limit: dates for other parts of the Act have since moved. Our explainer on the EU AI Act high-risk dates covers the later changes; confirm current dates on the Commission's page before relying on any of them.

6. A gold-medal-level result at the International Mathematical Olympiad

On July 21, 2025, Google DeepMind reported that an advanced version of Gemini Deep Think solved five of the six International Mathematical Olympiad problems, scoring 35 of 42 points, which is gold-medal level. It says the solutions were graded and certified by IMO coordinators using the same criteria as for students, and that the system worked in natural language within the 4.5-hour time limit, without translating problems into a formal language.

Why it matters: competition mathematics requires multi-step proofs, which makes it a more demanding test than short-answer questions. Limit: DeepMind's post states that the IMO confirmed the submitted answers were complete and correct, but its review did not extend to validating the system, its processes or the underlying model. It is also one vendor's report on one event, using a version of the model that was not generally available. See our guide to reading benchmark claims.

7. The US AI Action Plan

On July 23, 2025, the White House published Winning the AI Race: America's AI Action Plan. Its page says the plan follows a January executive order and identifies more than 90 federal policy actions, organised under three pillars: Accelerating Innovation, Building American AI Infrastructure, and Leading in International Diplomacy and Security.

Why it matters: it set out the US federal government's stated priorities for AI in one document, with infrastructure and export policy prominent. Limit: an action plan lists intentions. Which actions are carried out, funded or challenged has to be tracked separately, and we make no claim here about its results.

8. Genie 3: an interactive world model in a research preview

On August 5, 2025, Google DeepMind introduced Genie 3, which it calls a general-purpose world model able to generate interactive environments from a text prompt. The post claims navigation in real time at 24 frames per second at 720p, with environments staying largely consistent for several minutes. It lists limits too: a restricted action space, difficulty modelling several independent agents, imperfect geographic accuracy, and interaction lasting a few minutes, not hours. Access was limited to a small group of academics and creators.

Why it matters: world models are a route towards training and testing agents in generated environments. Limit: the performance details are the developer's claims about a research preview that the public could not use.

9. gpt-oss: open-weight models from OpenAI

In August 2025 OpenAI released open-weight models named gpt-oss. The model card for gpt-oss-120b on Hugging Face lists a permissive Apache 2.0 licence, 117 billion total parameters with 5.1 billion active, three adjustable reasoning levels, full chain-of-thought access that it says is not intended to be shown to end users, and agentic features such as function calling, web browsing and Python execution. The card shows no explicit release date; it lists the accompanying paper as published August 8, 2025 and the collection as updated August 7, 2025, so we say only 'August 2025'.

Why it matters: it added a large US lab to the group releasing weights people can run themselves. Limit: OpenAI's own site blocked our reader, so we rely on the model card. Weights under an Apache licence are not the same as open training data or a fully open process.

10. The Model Context Protocol moves to neutral governance

On December 9, 2025, Anthropic announced it was donating the Model Context Protocol (MCP), the standard it had introduced about a year earlier for connecting AI applications to tools and data, to the Agentic AI Foundation, a directed fund under the Linux Foundation. The page names Anthropic, Block and OpenAI as co-founders, lists goose and AGENTS.md among the founding projects, and says MCP's governance remains unchanged. MCP's own versioning page shows the protocol uses dated version identifiers, and its latest revision as of our reading carries a July 2026 date.

Why it matters: a connector standard is only useful if vendors trust it will not be steered by one company, and a neutral home is a common route to that. Limit: adoption figures in announcements come from the organisations involved and we did not verify them.

What do the ten have in common?

How can I check a claim like these myself?

What are the limitations of this list?

• The ten are our selection under the criteria above. Other reasonable lists would differ.
• We read primary pages with an automated reader on one day, October 18, 2026. Some vendors' own sites blocked it, so launches we could not read directly are not included.
• We did not test any model, product or chip, and we quote no benchmark scores.
• Some items have changed since 2025; where a source showed a later change we say so, but we did not audit every entry for later developments.
• If you find an error, tell us and we will record it on our corrections page.