Welcome back!

OpenAI just dropped GPT-4.1, and while it's not reaching #1 in benchmarks, it's a quiet coding powerhouse. With a 1 million-token context, blazing speed, and cheap pricing, this release is built for devs who need reliability. Fewer hallucinations, better code, faster responses. Let’s unpack…

Today’s Summary:

• 🚀 OpenAI launches GPT-4.1

• 🤯 DeepCoder-14B challenges top models

• 🎓 Claude becomes CS student favorite

• 🔚 GPT-4.5 to be removed from API

• 💰 Ilya Sutskever’s SSI hits $32B valuation with no product

• 🎀 ChatGPT action figure trend explodes in social media

• 🛠️ 2 new tools

TOP STORY

OpenAI launches GPT-4.1

The Summary: OpenAI has released GPT-4.1, mini, and nano, a new family of API-only models built for developers. All three models support 1M-token context windows (with no extra cost), faster response times, and better instruction following. The 4.1 series is designed to outperform GPT-4o in real-world dev tasks like coding, agents, and reasoning, while being cheaper to run.

Key details:

• API-only; won’t ship in ChatGPT

• SWE-Bench score: 55% vs 33% for GPT‑4o and 62% for Claude 3.7

• 1M token context with sub-5s latency on nano

• Reduced irrelevant code edits from 9% to 2%

• 26% cheaper than GPT-4o; nano runs at $0.12 per million tokens

• Already available in tools like VS Code; Windsurf offers 7 days free

Why it matters: With GPT‑4.1, OpenAI doesn’t seem focused on beating Gemini or Claude. Instead, it’s targeting some specific real-world pain points like slow coding responses, flaky agents, and forgetting long context. These models are built for developers shipping real products.

GPT‑4.1 Prompting Guide

OPEN SOURCE

DeepCoder 14B matches elite coding models

The Summary: DeepCoder-14B, a fully open-source 14B parameter model, reaches parity with OpenAI o3-mini. It scores 60.6% accuracy on LiveCodeBench and runs efficiently thanks to custom reinforcement learning and a novel training pipeline. Surprisingly, it also generalizes to math tasks, despite being trained only on code.

Key details:

• DeepCoder-14B hits 60.6% on LiveCodeBench, matching o3-mini

• Fine-tuned from DeepSeek-R1 by Together AI and Agentica

• Trained on 24K hand-filtered coding tasks

• Despite being a code-only model it scored 73.8% on AIME2024 math

Why it matters: DeepCoder gives o3-mini-level coding performance in an open-source model you can run, modify, and fine-tune. It handles long prompts and stays lightweight at just 14B parameters. If you’re building with LLMs, this one’s worth checking out.

Run it locally on LM Studio

ANTHROPIC

Claude emerges as the study buddy of choice for CS majors

The Summary: Anthropic studied 600,000 student chats with Claude to explore how college students use AI. Computer Science students dominate usage, turning to Claude for coding and learning complex concepts. The trends raise big questions about where AI fits in the future of education.

Key details:

• STEM students are early adopters of AI tools, with CS overrepresented

• Students from different fields engage with AI in different ways

• AI is also being used to outsource higher-order cognitive tasks

• Positive uses include explaining concepts and create study materials

• There’s a growing debate on how traditional homework should evolve as AI becomes more embedded in learning

Why it matters: AI is quickly becoming part of how students learn and study. As these tools enter daily academic workflows, it’s essential that they support, but not replace, real understanding. Anthropic’s data gives an early look into how education might adapt.

QUICK NEWS

Quick news

• GPT-4.5 will be removed from the API on 14 July 2025

• Ilya Sutskever’s SSI raises $2B at $32B valuation

• ChatGPT action figure visual trend is blowing up social media

TOOLS

🥇 New tools

• Grimo - Cursor for writing

• Vapi - Voice agents for developers