6 events
Anthropic drops into the terminal, OpenAI answers with pre-training
On 24 February, Anthropic released Claude 3.7 Sonnet, letting one model switch between answering immediately and thinking first. Alongside it came Claude Code, a research preview that writes code from the terminal.
On the 27th, OpenAI released GPT-4.5 as a research preview. Set apart from the models that spend time reasoning, it was positioned as a scaled-up pre-training run. Its price, among other things, made the difference in direction easy to read.
On the 10th, Anthropic launched the Anthropic Economic Index, saying it would report regularly on which occupations and tasks Claude was being used for.
On the 2nd, Karpathy floated "vibe coding" on X — his name for writing software by giving in to the vibes and letting AI write code you don't read. The post drew millions of views, and the phrase quickly took on a life of its own.
This block is written by the editors. It is kept separate from the sourced record below.
OpenAI released deep research, which hands an investigation over to ChatGPT. Given a prompt, it works through the web in multiple steps, reads hundreds of sources and returns a written report — work that would take a person hours, done in tens of minutes. It runs on a version of o3 tuned for web browsing and data analysis. ChatGPT Pro got it first with a limit of 100 queries a month, followed by Plus, Team and Enterprise.
Andrej Karpathy floated the phrase "vibe coding" in a post on X — his name for a way of building software where you give in to the vibes, forget the code exists, and accept the AI's suggestions without reading the diffs. The post drew millions of views, and the phrase spread rapidly, putting a name to a style of development already taking hold. In November, Collins Dictionary named it the 2025 word of the year. Japanese adopted "バイブコーディング" as-is, with explainers followed by articles questioning the quality and security of the resulting code.
Anthropic published the first report of the Anthropic Economic Index, an ongoing effort to track AI's effects on labor markets and the economy. The method analyzes roughly a million anonymized claude.ai conversations with a privacy-preserving tool called Clio and maps them onto the roughly 20,000 job tasks in the US Department of Labor's O*NET database. Usage clustered in software development and technical writing, and skewed toward tasks in mid-to-high wage occupations such as programmers and data scientists while being thinner at both the lowest- and highest-paid ends. Regularly publishing what its own service is actually used for drew attention as a form of transparency from an AI company.
The AI Action Summit, co-chaired by President Emmanuel Macron of France and Prime Minister Narendra Modi of India, ran for two days at the Grand Palais in Paris. Sixty-two countries, the African Union Commission and the European Union signed a joint statement on inclusive and sustainable AI for people and the planet, naming priorities such as closing digital divides, transparency and trustworthiness, effects on labour markets and environmental cost. The United States and the United Kingdom did not sign. US Vice President JD Vance argued in his opening remarks that excessive regulation would stifle innovation, while the UK cited national security concerns. With the summit's name having shifted from "safety" to "action", it was read as a change of course from Bletchley.
Anthropic released Claude 3.7 Sonnet, a hybrid reasoning model that switches between ordinary responses and "extended thinking" that shows its reasoning, with API users able to set a thinking budget of up to 128K tokens. Pricing held at $3 per million input tokens and $15 per million output tokens, thinking tokens included. Alongside it came Claude Code, a limited research preview of an agentic tool that takes on engineering work from the terminal — searching code, editing files, running tests, and driving GitHub and command-line utilities. Handing over a task at the terminal, rather than offering completions inside an editor, was an unfamiliar shape at the time.
OpenAI released GPT-4.5, which it described as its largest and most knowledgeable model to date, as a research preview. By scaling up unsupervised learning, the model was said to improve pattern recognition, breadth of knowledge, conversational naturalness and emotional intelligence (EQ), while reducing hallucinations. Unlike the reasoning-focused o-series, it followed a different track, and was initially offered to ChatGPT Pro users and API developers.