Skip to main content
Back

OpenAI Reveals GPT-4.5: Their Largest and Best Model for Chat Yet

LEARN MORE

February 28, 2025 7 min read

Hey,

Welcome to AlphaSignal – the most read newsletter by AI developers. 


We bring you the top 1% of news, papers, models, and repos, all summarized to keep you updated on the latest in AI.

IN TODAY'S SIGNAL

Read time: 7 min 02 sec

🎖️ Top News

  • OpenAI announces GPT-4.5, a model with expanded knowledge, better user intent understanding, and enhanced emotional intelligence.

📌  Activeloop


⚡️ Trending Signals


📝  Top Papers

🧠  Deep Dive

If you're enjoying AlphaSignal please forward this email to a colleague. 

It helps us keep this content free.

TOP NEWS

Chatbot

OpenAI unveils ChatGPT-4.5, enhancing human intent recognition and emotional intelligence

⇧ 92,583 Likes

What's New

OpenAI introduces GPT-4.5, a high-compute model with improved performance over GPT-4o. It achieves stronger results in math, science, multilingual understanding, and coding, while also demonstrating higher emotional intelligence and creativity. While GPT-4o remains more efficient, GPT-4.5 delivers higher accuracy in structured outputs and complex problem-solving.

Improved Capabilities

GPT-4.5 introduces optimizations in knowledge retrieval, multi-step execution, and structured outputs.

  • Supports function calling, structured outputs, system messages, and streaming in the API.

  • Uses image inputs for vision-based applications.

  • Improves agentic planning and execution for multi-step automation and coding workflows.

  • Enhances world knowledge interpretation with better accuracy and reduced hallucinations.

  • Demonstrates higher emotional intelligence and creativity, improving writing, communication, and brainstorming tasks.

Performance on Benchmarks

GPT-4.5 scores higher than previous models across multiple technical and academic benchmarks.

  • 38.0% on SWE-Bench Verified (coding) vs. 30.7% for GPT-4o.

  • 71.4% on GPQA (science) vs. 53.6% for GPT-4o.

  • 36.7% on AIME ‘24 (math) vs. 9.3% for GPT-4o.

  • 85.1% on MMMLU (multilingual) vs. 81.5% for GPT-4o.

  • 74.4% on MMMU (multimodal) vs. 69.1% for GPT-4o.

Higher Emotional Intelligence and Creativity

GPT-4.5 improves on human-like communication and creative problem-solving.

  • Better recognition of nuanced emotions in text-based interactions.

  • More coherent and expressive storytelling and content generation.

  • Improved adaptability in conversational tone and response style.

Usecase

  • Writing assistance and communication tools that benefit from improved "EQ".

  • Coaching and brainstorming applications requiring nuanced responses.

  • Multi-step coding workflows and complex task automation.

Availability

GPT-4.5 is accessible  but requires more compute resources than GPT-4o.

  • Available in the Chat Completions API, Assistants API, and Batch API.

  • More expensive than GPT-4o due to increased compute demands.

  • OpenAI is evaluating its long-term availability based on demand and cost-effectiveness.

    Community Feedback


    Sam Altman

    "this isn’t a reasoning model and won’t crush benchmarks. it’s a different kind of intelligence and there’s a magic to it i haven’t felt before. really excited for people to try it!"

    Cursor

    "GPT 4.5 in Cursor! We've found it surprisingly effective in cases where all other models fail."

    Apollo

    "Given its purpose -non reasoning model upgrade over gpt4o- it should have support for multimodal features from the start. That is definitely a step back"

    Learn more

    Run Multi-Modal Deep Research on Your Data 

    Get answers from your private and public data.

    Activeloop AI Knowledge Agents:

    • Search billions of data points across images, PDFs, text, tables, and more in S3, GCP, or Dropbox.

    • Use vision-language models like CoLPaLI for analysis.

    • Extract context, perform deep reasoning, and return accurate answers with your chosen LLM.

    • Improve responses over time based on your queries.

    • Handle large-scale queries efficiently.

    • Support multi-modal data processing.

    • Connect seamlessly to your stored data.


    Used by companies like Bayer Radiology, Activeloop simplifies complex research in finance, legal, life sciences, healthcare, insurance, and logistics.

    Use promo code ALPHA to get a 1-month free trial now.

    Try now

    partner with us →

    TRENDING SIGNALS

    LLM Usecase

    Andrej Karpathy releases a new video, exploring LLM interactions and tools like Custom GPTs, NotebookLM, and Sora

    ⇧ 16,395 Likes

    Document Processing

    Andrew Ng introduces Agentic Document Extraction, a new approach for PDF parsing for better RAG and multimodal document analysis

    ⇧ 18,485 Likes

    AI Optimization

    DeepSeek presents 3FS, which optimizes KVCache, dataset loading, and checkpointing for large-scale AI workloads

    ⇧ 5,693 Likes

    Small Language Models

    Microsoft unveils Phi-4, an open-source series of compact multimodal model excelling in text, vision, and speech tasks

    ⇧ 4,729 Likes

    LLM

    Stanford professor led Inception Labs launches Mercury, a diffusion-based LLM, delivering 10x faster text generation and cheaper

    ⇧ 7,385 Like

    TOP PAPERS

    AI Model

    Claude’s extended thinking

    ⇧ 2,495 Likes

    Problem

    AI models struggle with complex, multi-step reasoning and maintaining focus over extended interactions. Traditional models often limit their computational effort per query, affecting performance on tasks requiring deep thought and iterative problem-solving.

    Solution
    Claude 3.7 Sonnet introduces "extended thinking mode," allowing it to allocate more computation for difficult problems. It also improves agent capabilities, enabling iterative decision-making in interactive environments. The model’s visible thought process enhances transparency, and new test-time compute scaling methods boost accuracy on scientific and reasoning benchmarks.

    Results
    Claude 3.7 Sonnet achieves higher accuracy on math, science, and logic tasks by leveraging extended and parallel test-time compute. It outperforms previous versions in multimodal AI agent evaluations like OSWorld and can sustain complex tasks, such as playing Pokémon Red. While improvements in transparency aid trust, challenges remain in ensuring faithfulness and preventing adversarial misuse.

    Reasoning Models

    Chain of Draft: Thinking Faster by Writing Less

    ⇧ 2,045 Likes

    Problem

    Large Language Models (LLMs) like Chain of Thought (CoT) excel at complex reasoning tasks by breaking problems into verbose, step-by-step processes. However, this approach results in high computational costs, latency, and excessive token usage.

    Solution
    The paper proposes Chain of Draft (CoD), a novel reasoning strategy inspired by human cognition. CoD encourages LLMs to generate concise, dense intermediate outputs, focusing on essential insights without unnecessary elaboration, reducing verbosity and enhancing efficiency.

    Results
    Experiments show that CoD achieves similar or better accuracy compared to CoT while using only 7.6% of the tokens, significantly lowering costs and latency across various reasoning tasks. This minimalist approach improves LLM performance, making it more practical for real-world applications.

    LLM Evaluation

    Can LLMs Generate Novel Research Ideas?

    ⇧ 5,249 Likes

    Problem

    While LLMs have shown promise in accelerating scientific tasks, it remains unclear whether they can generate novel, expert-level research ideas. This study aims to evaluate LLMs' ideation capabilities in comparison to human experts.

    Solution
    The study recruited over 100 NLP researchers to generate novel research ideas and perform blind reviews of both LLM and human-generated ideas. A rigorous experimental design controlled for confounders and compared the novelty and feasibility of ideas.

    Results
    LLM-generated ideas were judged as more novel than human-generated ideas (p < 0.05), though slightly weaker in feasibility. Despite excitement over LLM performance, the study revealed challenges, including lack of diversity in generated ideas and failures in LLM self-evaluation, pointing to open issues in developing research agents.

    DEEP DIVE

    AI Model

    Andrej Karpathy Analyzes the Advancements in GPT-4.5

    ⇧ 7,148 Likes

    This post analyzes the release of GPT-4.5, highlighting the subtle improvements from scaling pretraining compute. It compares GPT-4.5 to earlier models, discussing its enhancements in creativity, world knowledge, and humor, with a focus on tasks that require emotional intelligence rather than reasoning.

    Key topics covered:

    • Scaling pretraining compute and model performance

    • GPT-4.5's improvements in language creativity and world knowledge

    • The difference in reasoning tasks versus emotional intelligence tasks

    • Interactive comparison between GPT-4 and GPT-4.5 using polls and prompts

    • Community feedback and reactions on GPT-4.5's performance
    Learn more

    At Alpha Signal, our mission is to build a sharp, engaged community focused on AI, machine learning, and cutting-edge language models, helping over 200,000 developers stay informed and ahead. We’re passionate about curating the best in AI, from top research and trending technical blogs to expert insights and tailored job opportunities. We keep you connected to the breakthroughs and discussions that matter, so you can stay in the loop without endless searching. We also work closely with partners who value the future of AI, including employers and advertisers who want to reach an audience as passionate about AI as we are.


    Our partnerships are based on shared values of ethics, responsibility, and a commitment to building a better world through technology.Privacy is a priority at Alpha Signal. Our Privacy Policy clearly explains how we collect, store, and use your personal and non-personal information. By using our website, you accept these terms, which you can review on our website. This policy applies across all Alpha Signal pages, outlining your rights and how to contact us if you want to adjust the use of your information. We’re based in the United States. By using our site, you agree to be governed by U.S. laws.

    Looking to promote your company, product, service, or event to 200,000+ AI developers? Let's work together.

    Work with us