Welcome back!

OpenAI just dropped GPT-4.5 "Orion", their most advanced pre-trained model, promising more human-like understanding and fewer hallucinations. Early users say it feels like more chatting with a person rather than a chatbot. However, it requires massive computing resources. Let’s unpack...

Today’s Summary:

• 🚀 OpenAI launches GPT-4.5 Orion

• 💻 Google offers free Gemini Code Assist

• 🤖 Amazon revamps Alexa with Agentic AI

• 💬 ElevenLabs unveils Scribe speech-to-text

• 🔎 OpenAI Deep Research expands access

• 🔄 DeepSeek teases upcoming R2 model

• 🛠️ 2 new tools

TOP STORY

OpenAI Releases GPT-4.5 with human-like understanding

The Summary: OpenAI has released GPT-4.5 “Orion”, their largest and most advanced pre-trained model, which seems to grasp human intent better than previous versions. Early users report that GPT-4.5 feels more like talking to a person rather than an AI, with improvements in factual accuracy and reduced hallucinations. Due to high computing resource demands, the rollout is progressive, beginning today for Pro users before expanding to Plus users next week.

Key details:

• GPT-4.5 is OpenAI’s largest pre-trained LLM with increased knowledge, intuition, emotional intelligence, improved copywriting, and reduced hallucinations

• Andrej Karpathy views GPT-4.5 as a measure of what kind of improvements we can get from scaling pretraining compute 10x

• GPT-4.5 scores 62.5% accuracy on SimpleQA tests versus 38.2% for GPT-4o, with 37.1% hallucination rate compared to 61.8%

• Integrates with file and image upload, canvas, and search

• “Reasoning models” such as OpenAI o3-mini, DeepSeek-R1 and Claude 3.7 Sonnet, remain superior in math and science tasks

• OpenAI admits the model is “giant and expensive” and “not a replacement” for GPT-4o, with uncertain long-term API availability due to massive compute requirements, while the API is 25x more costly than GPT-4o

Why it matters: This release shows that while GPT-4.5 makes meaningful progress, traditional pre-training methods are reaching a point of diminishing returns with only incremental improvements. The industry now faces key questions about how to move AI forward, with OpenAI planning to combine GPT-4.5 with the o3 reasoning model in their next GPT-5 release, expected later this year.

System card

GOOGLE

Gemini Code Assist is now free

The Summary: Google has released a free version of Gemini Code Assist for individual developers, offering a huge 180,000/month code completions. The AI assistant integrates directly with popular IDEs like VS Code and JetBrains, letting developers generate, explain, and improve code through natural language prompts. Google also launched Gemini Code Assist for GitHub, which automatically reviews code in both public and private repositories.

Key details:

• Offers 180,000/month completions vs GitHub Copilot's 2,000 limit

• Features 240 daily chat requests and 128,000-token context

• Requires only a Gmail account with no credit card needed

• Automatically suggests code changes and detects bugs

Why it matters: The move democratizes access to free AI coding tools for students and hobbyists. As the AI coding battle heats up with recent entries from Anthropic (Claude Code) and Amazon (Q Developer), Google's generous free tier may help it gain market share while nurturing future enterprise customers.

FAQ

AMAZON

Amazon transforms Alexa with Agentic AI

The Summary: Amazon has unveiled Alexa+, a generative AI-powered assistant that moves beyond voice commands to complete complex tasks autonomously. The new assistant can book restaurants, make purchases, navigate websites, and remember your preferences without supervision. Alexa+ uses a mix of AI models and "experts" to handle specialized tasks.

Key details:

• Alexa+ works across 600 million devices and will be free for Prime members or $19.99 monthly for others

• Can autonomously navigate websites to complete tasks

• The system lets upload documents, emails, and photos

• Amazon employs a "model-agnostic" approach, using both their own Nova models and Anthropic's Claude

• Early access begins next month for US users

Why it matters: Alexa+ is bringing advanced genAI to 200 million Prime members, a scale that could redefine consumer expectations for digital assistants. With its shift to agentic autonomous task completion, Amazon is positioning Alexa+ to become an advanced household assistant.

QUICK NEWS

Quick news

• ElevenLabs launches Scribe, accurate speech-to-text model

• OpenAI Deep Research now available to ChatGPT Plus, Team, Edu, and Enterprise users, including in the EU

• DeepSeek to launch R2 model before May

TOOLS

🥇 New tools

• TestAI - AI voice agent testing & analytics platform

• Deep Review - AI co-scientist for literature reviews