AI Roundup February 2025: GPT-4.5, Claude 3.7 + Claude Code, and Grok 3 Crash the Party

TL;DR

February was a frontier-model pileup. Anthropic shipped Claude 3.7 Sonnet, its first hybrid reasoning model, and quietly dropped the tool that would define the year: Claude Code. xAI launched Grok 3 with a Super Bowl-sized ego. OpenAI ended the month with GPT-4.5, its biggest and most expensive model ever, and a muted reception. Meanwhile, Paris hosted a global AI summit that the US and UK pointedly refused to sign.

Anthropic Ships Claude 3.7 Sonnet and Claude Code on the Same Day

On February 24, Anthropic released Claude 3.7 Sonnet, its first hybrid reasoning model. The pitch: one model, two modes. You get fast standard answers, or you flip on extended thinking and let it burn tokens working through hard problems step by step. No separate reasoning model to juggle, no model picker roulette. For builders, the killer detail was control over the thinking budget through the API, so you decide how much compute a hard problem gets.

The same drop included a research preview of Claude Code, a command-line coding agent that lives in your terminal and edits your repo directly. At the time it read like a neat side project. In hindsight it was the start of the agentic-coding era, and the tool a lot of you now leave running all day.

One model that answers instantly or thinks hard on demand, plus an agent that actually touches your filesystem. That combination is the template everyone copied.

xAI Launches Grok 3 With Maximum Confidence

On February 17, Elon Musk's xAI unveiled Grok 3 in a livestreamed event, calling it the smartest AI on Earth and waving around benchmark charts to prove it. It launched with a Think mode for reasoning and a DeepSearch agent for live web research, initially gated behind X Premium+ and a new SuperGrok tier.

The interesting story here is not the leaderboard jockeying, it is the compute. Grok 3 was trained on the Colossus cluster in Memphis, a build-out xAI stood up at a pace that made the rest of the industry blink. Whatever you think of the bravado, going from a standing start to a credibly frontier-class model this fast is the part worth respecting.

OpenAI Closes the Month With GPT-4.5, Its Biggest and Priciest Model

On February 27, OpenAI released GPT-4.5, codenamed Orion, as a research preview for Pro and Plus users. It was explicitly not a reasoning model. The bet was the old one: scale up pre-training and post-training and get a model with broader knowledge, better instinct, and fewer hallucinations. OpenAI cited a hallucination rate of 37.1 percent versus 59.8 percent for GPT-4o, and pitched it as warmer and more natural to talk to.

The reception was muted, and the reason was the price tag. Through the API, GPT-4.5 ran 75 dollars per million input tokens and 150 dollars per million output tokens. That is roughly 30 times the input cost of GPT-4o. For a non-reasoning model arriving the same week as cheaper, smarter, thinking models, the math was brutal. OpenAI began winding GPT-4.5 down from the API just weeks later. Read it as the clearest signal yet that pure pre-training scaling had hit a wall of diminishing returns.

Why builders should care

• Reasoning won the month. The cheaper hybrid and reasoning models stole the spotlight from the giant base model.
• Price is product. A model you cannot afford to call in a loop is not a model you will build on.

Paris Hosts an AI Summit the US and UK Won't Sign

On February 10 and 11, France hosted the AI Action Summit at the Grand Palais, with representatives from more than 100 countries. The final declaration called for AI that is open, inclusive, transparent, ethical, and safe. Sixty-one nations and bodies signed, including France, China, India, and the EU.

The US and UK did not. US Vice President JD Vance used his appearance to argue that excessive regulation would smother American AI, while the UK cited concerns about national security and global governance clarity. The split mattered: it marked a clear pivot from the safety-first tone of earlier summits toward an accelerate-and-compete posture from Washington. If you build on this stuff, that shift in the regulatory wind shapes everything downstream.

Key Takeaways

• The frontier got crowded. Three labs shipped flagship models in a single month, and no one vendor owns the lead anymore.
• Reasoning beat raw scale. GPT-4.5's expensive, non-reasoning approach landed with a thud the same week hybrid and thinking models shined.
• Claude Code was the sleeper. The terminal agent that looked minor in February turned out to be the most consequential thing Anthropic shipped that month.
• Control is a feature. Tunable thinking budgets and mode switches are now table stakes, not novelties.
• Policy split in two. The Paris summit made the US accelerationist stance official, and that posture sets the backdrop for everything you deploy.