January 2025: The DeepSeek Shock

On January 20, 2025, a Chinese AI lab called DeepSeek released an open-weight reasoning model called R1. Within seven days, it had topped the Apple App Store charts in both the United States and China and accumulated over 100 million users.

The numbers that stunned the industry: DeepSeek claimed R1 was trained for approximately $6 million — a figure that has not been independently verified. OpenAI's GPT-4 was widely reported to cost somewhere in the range of tens to hundreds of millions of dollars to train. If the efficiency gap is even close to accurate, it suggests the assumption that frontier AI requires massive compute investment may have been overstated. If the efficiency claim holds up under scrutiny, it suggests the assumption that frontier AI requires tens of billions in compute investment may have been wrong.

Markets reacted immediately. Nvidia's stock fell 18% in a single day — a loss widely reported at approximately $593 billion in market capitalization — one of the largest single-day market cap drops in US stock market history.

The technical significance: DeepSeek R1 used a technique called Mixture of Experts (MoE) more aggressively than Western labs, activating only a fraction of its parameters per query. Combined with innovations in training data efficiency and reinforcement learning, it achieved performance comparable to OpenAI's o1 at a fraction of the cost. The implications for the "who controls compute controls AI" assumption are still being worked out.

February 2025: Anthropic's Extended Thinking

On February 24, 2025, Anthropic released Claude 3.7 Sonnet with a new capability called Extended Thinking — a visible chain-of-thought mode that allows the model to reason through problems before generating a response. Users can watch the thinking process unfold in real time.

On GPQA Diamond — a benchmark built from questions designed by PhD scientists in physics, chemistry, and biology, where human domain experts average around 65% — Claude 3.7 Sonnet scored 84.8%. That placed it ahead of all other publicly available models at the time of release.

Anthropic simultaneously launched Claude Code, an agentic programming tool designed to handle complex, multi-step coding tasks asynchronously. This marked Anthropic's first serious move into the "AI agent" product space that would dominate the rest of the year.

March 2025: Gemini 2.5 Pro Takes the Lead

Google released Gemini 2.5 Pro Experimental in March 2025. Within days of release, it claimed the top spot on the LMSYS Chatbot Arena — the largest public head-to-head AI evaluation platform — beating GPT-4.5 by approximately 40 Elo points. That margin is considered substantial in a leaderboard where differences of 10–15 points typically signal meaningful capability gaps.

Gemini 2.5 Pro introduced a "Deep Think" reasoning mode and a 1 million token context window. On Poe, the AI aggregator platform, it captured approximately 30% of all reasoning query volume within six weeks of launch.

March also marked a turning point for AI interoperability. Anthropic's Model Context Protocol (MCP) — an open standard for connecting AI models to external tools — gained mainstream adoption when OpenAI's ChatGPT announced support for the protocol. Google confirmed support in April. MCP is now effectively the industry standard for AI tool integration.

April–May 2025: Meta and OpenAI Expand the Field

Meta released Llama 4 in April 2025, with two variants: Scout (for efficiency) and Maverick (for reasoning). Both were open-weight models, meaning the parameters were publicly downloadable. Llama 4 Maverick competed with GPT-4.5 on several benchmarks and represented the most capable open-weight model released to that point.

OpenAI followed in April with o3 and o4-mini — the next generation of its reasoning model series — along with GPT-4.1, an update focused on instruction following and reduced latency. The pace of releases was accelerating to the point where the industry had largely stopped treating individual model launches as landmark events and started treating them as routine updates.

In May 2025, Anthropic released Claude 4, including Opus 4.5 and Sonnet 4.5. The Opus variant was positioned as a document analysis and enterprise research model. It significantly improved over its predecessor on long-context tasks and multi-step reasoning.

July 2025: AI Wins the IMO

In July 2025, both OpenAI's reasoning model and Google DeepMind's Gemini Deep Think achieved gold-medal-equivalent performance at the International Mathematical Olympiad — independently, in the same competition cycle. Gemini Deep Think scored at a gold-medal-equivalent level, completing all problems in natural language within the standard time limit. Google DeepMind reported the score as 35 points out of a possible 42.

For context: the IMO is the most prestigious high school mathematics competition in the world. Human gold medalists are among the most mathematically gifted individuals alive. The fact that two separate AI systems achieved this standard in the same year, using fundamentally different architectures, suggests this was not a lucky result.

The same month, both systems also achieved top placements in the International Collegiate Programming Contest (ICPC). These results marked the moment when AI crossed the threshold from "competitive with strong human mathematicians" to "competitive with the very best."

August 2025: GPT-5 and the EU AI Act

OpenAI released GPT-5 on August 7, 2025. The model introduced dynamic "thinking modes" — allowing users to select between fast responses and extended reasoning — and was reported by OpenAI to have a significantly reduced hallucination rate compared to GPT-4. It handled text, images, and structured data natively.

The same month carried regulatory significance: August 2, 2025 marked the date when provisions governing General Purpose AI (GPAI) models under the EU AI Act formally came into effect. This is the world's first comprehensive AI law. Under the GPAI rules, providers of high-capability foundation models must conduct adversarial testing before deployment, maintain technical documentation, comply with EU copyright law, and publish summaries of training data.

The EU's enforcement is already active. In Q1 2026, EU member states were reported to have issued dozens of fines totaling hundreds of millions of euros, primarily for GPAI non-compliance. Ireland, which hosts the European headquarters of most major US tech companies, was reported to have handled the majority of cases.

September–November 2025: The Year-End Sprint

DeepSeek published a research paper in September 2025 that appeared on the cover of Nature — a peer-reviewed scientific journal whose cover placement is considered one of the highest marks of research significance. It was the first time an AI company's technical publication had received that placement.

The year's final months produced a cascade of flagship model releases:

November 12: OpenAI released GPT-5.1, with improvements in latency, tool use, and instruction following.

November 17: Grok 4.1 from xAI was released.

November 18: Google released Gemini 3 Pro — the first Google model to claim the top position on the Artificial Analysis Intelligence Index and the first model from any lab to exceed 1,500 Elo on the LMSYS Chatbot Arena.

December 11: OpenAI released GPT-5.2, which was reported to be among the first models to score above 90% on ARC-AGI-1 — a benchmark specifically designed to test novel reasoning rather than pattern recall. It also achieved a perfect score on AIME 2025. Alongside the model, OpenAI launched Codex, an autonomous programming agent designed to handle entire engineering tasks with minimal human oversight.

2026: The Frontier Continues Moving

The first months of 2026 accelerated rather than slowed the pace of development.

Dario Amodei, Anthropic's CEO, told the World Economic Forum in Davos in January 2026 that AGI-level systems were likely "within a few years" — pointing to 2027 as a plausible horizon. Shane Legg, DeepMind's co-founder, gave a 50% probability of "Minimal AGI" by 2028. These are not fringe predictions; they come from the people building the systems.

In April 2026, Anthropic released Claude Mythos 5 — a 10-trillion-parameter model with a focus on cybersecurity and advanced coding. Google released Gemini 3.1 with real-time voice and image analysis capabilities and a Flash-Lite variant running at 2.5 times the speed of its predecessor.

On April 23, 2026, OpenAI released GPT-5.5, internally codenamed "Spud" — the first fully retrained base model since GPT-4.5. It scored 85% on ARC-AGI-2 (a harder successor benchmark) and was reported to have topped the Artificial Analysis Intelligence Index. On OSWorld-Verified, a benchmark testing AI's ability to autonomously operate real computer environments, it scored in the high-70% range according to published reports.

The White House released a National Policy Framework for Artificial Intelligence on March 20, 2026, offering legislative recommendations for unified governance. No comprehensive federal AI law exists yet in the United States, while California, Colorado, New York, Illinois, and Utah have each enacted or proposed their own legislation — creating a fragmented regulatory environment that the tech industry has argued makes compliance planning extremely difficult.

The Through-Line

Across 18 months, a few patterns are clear.

First, the efficiency story changed. DeepSeek demonstrated that state-of-the-art performance does not require state-of-the-art compute budgets. This has implications for who can build frontier AI — not just the three or four US labs with billion-dollar infrastructure, but smaller teams with access to more efficient training techniques.

Second, reasoning became the dominant axis of competition. The shift from "what can the model output" to "how well can it think through hard problems" defines the 2025–2026 period. Extended thinking, chain-of-thought, and reinforcement learning from human feedback combined to produce the IMO and ARC-AGI results.

Third, agentic AI moved from research to product. Claude Code, OpenAI Codex, and Google Jules are not research prototypes — they are deployed tools that engineering teams are using today. The question for 2026 and beyond is not whether AI can do complex tasks but how much human oversight those tasks actually require.

Fourth, regulation arrived. The EU AI Act is generating real fines. State-level laws in the US are proliferating. China's amended Cybersecurity Law is in effect. The governance layer is catching up to the capability layer, though exactly how it will shape development over the next few years remains genuinely uncertain.

What is not uncertain: the rate of change. Whatever the state of AI looked like when you last checked, it has almost certainly changed since.