September 2025 was one of the most consequential months in AI history. In just 30 days, Nvidia announced plans to invest up to $100 billion in OpenAI, Anthropic raised $13 billion and settled a landmark $1.5 billion copyright case, OpenAI released GPT-5 Codex and the Sora 2 video app, Anthropic shipped Claude Sonnet 4.5, California signed the first major frontier AI safety law in the US, and Albania appointed the world’s first AI government minister.
Looking back from mid 2026, that month reads like a table of contents for everything that followed. This recap covers each major story and, more usefully, what it turned into since. If you want a quick verdict: September 2025 was the month AI’s money, models, and laws all shifted gears at once.
The Money Moves: Deals That Redefined Scale
Nvidia and OpenAI announced a deal of up to $100 billion. On September 22, the two companies signed a letter of intent for Nvidia to invest up to $100 billion in OpenAI as OpenAI builds out at least 10 gigawatts of Nvidia powered data centers. The number stunned even a market used to huge AI figures. It also ignited the “circular financing” debate, the concern that chipmakers funding their own customers inflates the ecosystem, an argument that has only grown louder since.
The Stargate project expanded. Days later, OpenAI, Oracle, and SoftBank announced five new US data center sites for the Stargate buildout, pushing the project toward its multi hundred billion dollar vision. The AI infrastructure race that dominates 2026 headlines was locked in during these weeks.
Anthropic raised $13 billion at a $183 billion valuation. At the start of the month, Anthropic closed one of the largest funding rounds in tech history. In hindsight it was a bargain checkpoint: by 2026, reports placed Anthropic’s valuation near the trillion dollar range as it filed for an IPO, alongside OpenAI’s own listing plans.
Nvidia also put $5 billion into Intel, a once unthinkable alliance between old rivals, signaling how completely AI demand had redrawn the chip industry’s map.
The Model Releases: A Big Month for Builders
OpenAI released GPT-5 Codex. Mid September brought a version of GPT-5 tuned specifically for agentic coding, able to work independently on long programming tasks inside the Codex product. It marked the moment coding agents became a headline product category rather than an experiment, a shift that reshaped developer workflows and pricing models, which we track in our guide to OpenAI Codex pricing.
Anthropic released Claude Sonnet 4.5. On September 29, Anthropic shipped what it called its best coding model yet, with the standout claim that it could sustain focus on complex, multi step tasks for many hours at a stretch. That endurance framing, models measured by how long they can work alone, became a defining benchmark across the industry in the year since.
xAI pushed speed and price. Elon Musk’s xAI released Grok 4 Fast in the same window, competing on efficiency, an early sign of the cost pressure that now dominates enterprise AI buying decisions.
The Product Launches: AI Went Consumer
OpenAI launched Sora 2 and a social video app. On September 30, OpenAI released its upgraded video model along with an invite only social app where every clip in the feed is AI generated, complete with a feature letting users insert their own likeness into scenes. It triggered immediate debates about deepfakes and copyrighted characters. The hindsight twist: by 2026, OpenAI shut the Sora app down, making it one of the era’s most instructive product experiments in how fast AI consumer bets can rise and fold.
ChatGPT got proactive with Pulse. In late September, OpenAI previewed Pulse, which researches overnight and delivers personalized morning briefings without being asked. It was an early step in the industry wide shift from chatbots that answer to assistants that act, the same shift driving the AI agents boom.
Meta launched Vibes, its own feed of AI generated short videos, and Google pushed Gemini directly into Chrome, embedding an assistant into the world’s most used browser. Every major platform, in the same month, bet that AI belongs inside the product rather than beside it.
The Rules Arrived: Law and Government Firsts
California signed SB 53, the first major US frontier AI law. On September 29, Governor Newsom signed the Transparency in Frontier Artificial Intelligence Act, requiring the largest AI developers to publish safety frameworks and report critical incidents. It became the template other states studied, and by 2026 the regulatory conversation it started had reached federal level talks on frontier model standards.
Anthropic settled with authors for $1.5 billion. In the first week of September, Anthropic agreed to pay authors at least $1.5 billion over pirated books used in training data, the largest copyright settlement of its kind. It put a real price tag on training data questions and pushed the whole industry toward licensing deals.
Albania appointed the world’s first AI minister. The country named Diella, an AI system, as a government minister responsible for public procurement, with the goal of corruption free tendering. Symbolic or not, a government formally handing a portfolio to an AI made September 2025 a genuine historical first.
Microsoft and OpenAI redefined their relationship. The two signed a preliminary agreement on restructuring their partnership, the deal that cleared the path for OpenAI’s corporate reorganization and, eventually, its move toward public markets.
What September 2025 Looks Like From 2026
Three threads from that month explain today’s AI landscape.
First, the infrastructure bets got locked in. The Nvidia OpenAI deal and Stargate expansion committed the industry to compute spending so large that questions about returns on it now dominate financial coverage.
Second, agents replaced chatbots as the frontier. GPT-5 Codex, Sonnet 4.5’s endurance claims, and Pulse all pointed the same direction: AI that works on its own. That is now the standard businesses build toward, and it is the core of what we help companies deploy through AI agent development and AI automation and workflow integration.
Third, the accountability era began. A billion dollar copyright settlement and the first frontier AI law in the same month ended the industry’s rules free adolescence.
For businesses, the practical lesson from that month still holds: the technology moves faster than most adoption plans, so strategy matters more than tool chasing. Sorting durable shifts from hype is exactly what our AI consulting and strategy work exists for.
Frequently Asked Questions
1. What was the biggest AI news in September 2025?
The Nvidia and OpenAI announcement of up to $100 billion in planned investment was the largest story, alongside Anthropic’s $13 billion raise, the $1.5 billion authors settlement, and major model releases including GPT-5 Codex, Claude Sonnet 4.5, and Sora 2.
2. What AI models were released in September 2025?
The month’s key releases were OpenAI’s GPT-5 Codex for agentic coding, Anthropic’s Claude Sonnet 4.5, xAI’s Grok 4 Fast, and OpenAI’s Sora 2 video model with its companion social app.
3. What was California’s SB 53?
Signed on September 29, 2025, it was the first major US law targeting frontier AI, requiring the largest developers to publish safety frameworks and report critical safety incidents. It became a model for wider AI regulation efforts.
4. Did Albania really appoint an AI minister?
Yes. In September 2025, Albania named an AI system called Diella as the minister responsible for public procurement, the first case of a government formally assigning a ministerial role to an AI.
5. Why does September 2025 still matter?
Because its events set the direction the industry followed: massive infrastructure spending, the shift from chatbots to autonomous agents, and the start of real AI regulation and copyright accountability.
