Anthropic quietly dropped Claude 4 Sonnet — and it's actually great
No trailer, no livestream, no long CEO tweet. Anthropic just dropped Claude 4 Sonnet straight into the API, and even trimmed the price a bit. I've been running it all day—coding and long-form reasoning are a solid step up from 3.5.
💡 What You Will Learn
No trailer, no livestream, no long CEO tweet. Anthropic just dropped Claude 4 Sonnet straight into the API, and even trimmed the price a bit. I've been running it all day—coding and long-form reasonin
📜 Table of Contents
Let me start with the conclusion: Claude 4 Sonnet is currently one of the most cost-effective models out there, but how it was released is more worth discussing than the model itself.
The Model Itself
I ran it through a few of my own scenarios:
Coding: I gave it a complex frontend state management problem. 3.5 Sonnet got stuck on handling intermediate states, while 4 Sonnet directly produced a complete solution with reducer, context, and localStorage persistence. It worked on the first try—no debugging needed.
Long-form reasoning: I handed it a 15-page paper to summarize and identify logical flaws. It caught the author's bias in sample selection—something I only noticed after reading it three times.
Multi-turn conversation: I asked 20 follow-up questions in a row, and it maintained consistency throughout without the "zoning out" issue 3.5 had in long conversations.
On pricing, input/output costs are roughly 15% cheaper than 3.5.
What's More Worth Discussing: The Release Approach
OpenAI rolls out a full week of marketing blitz for every new model. Anthropic just changed the model name in the docs and shipped it.
No announcement, no benchmark press releases, no "we're redefining artificial intelligence" rhetoric.
I suspect two reasons: 1. They position their product as a tool, not an event. Tool updates don't need launch events—just like Photoshop doesn't need Tim Cook on stage for a new version. 2. They know the benchmark game has lost all credibility. Instead of burning energy climbing leaderboards, they'd rather let users feel the difference through real usage.
Back in 2024, these two reasons might have been dismissed as "bad marketing." In 2026, they've become a form of confidence.
When the whole industry has learned to mask incremental progress with fanfare, silence becomes the most weighty statement.
Written by our editorial team; tools listed here are tested or verified against public sources. Links point to official sites or GitHub repos for reference only — no paid placements.
