In partnership with

Good morning. It’s Friday, March 7th.

Today in tech history: 99 years ago today, The first successful transatlantic radio-telephone conversation took place between New York and London.

In today’s email:

• OpenAI’s Rumored $20k Agents

• Google’s AI Search

• China’s Quantum Computer

• 2 New AI Tools

• Latest AI Research Papers

You read. We listen. Let us know what you think by replying to this email.

OpenAI Rumored to Release $20K/mo ‘AI agents’

OpenAI is rolling out AI agents priced between $2,000 and $20,000 per month, targeting enterprise and research clients, The Information reports. The company expects these tools to eventually make up 20-25% of its revenue, with SoftBank backing the effort with a $3 billion investment for 2025.

Offerings range from a $2,000 ChatGPT agent sorting sales leads to a $10,000 coding assistant for senior engineers. At the top tier, a $20,000 model is built for PhD-level research. With OpenAI burning through an estimated $5 billion last year, these high-margin products are a calculated bet to balance the books.

Simultaneously, GPT-4.5 is rolling out to ChatGPT Plus users, refining fluency and accuracy but falling short in reasoning. CEO Sam Altman admitted it wasn’t designed for benchmark supremacy, trailing models like Claude 3.7 and DeepSeek AI.

Despite expanded compute training, the model costs $150 per million tokens—15 times that of GPT-4o—raising questions about adoption. OpenAI is staggering access over three days due to GPU constraints, with full support for ChatGPT tools and APIs.

Expanding its AI footprint, OpenAI has also introduced ChatGPT for macOS with IDE integration, enabling direct code edits and automating routine development tasks for Plus, Pro, and Team subscribers.

However, OpenAI is also recalibrating its AGI stance. In a recent blog post confirms OpenAI no longer sees AGI as a dramatic leap forward but a slow, iterative climb. This shift has shaped its approach to safety, focusing on real-world testing over theoretical doomsday planning. Internal discussions highlight AGI’s financial stakes, with Microsoft defining it as AI that outperforms humans in key economic tasks and generates at least $100 billion in profit.

This repositioning has drawn criticism. Former policy researcher Miles Brundage has accused OpenAI of rewriting its safety history. OpenAI now frames AGI as a gradual progression rather than a sudden breakthrough, arguing that excessive caution—such as withholding GPT-2 in 2019—was unwarranted.

Brundage refutes this, asserting that GPT-2’s phased release aligned with OpenAI’s current iterative deployment philosophy. His criticism underscores concerns that OpenAI is downplaying safety risks to accelerate product releases, a shift that coincides with mounting financial pressure and competition. Read more.

Google launches ‘AI Mode’ in Search for complex, multi-step queries

Google is launching AI Mode, an experimental Search feature powered by Gemini 2.0. It runs parallel searches across web content, Knowledge Graph data, and live shopping feeds to deliver structured, interactive responses.

Unlike traditional search, AI Mode supports multi-step queries and real-time comparisons, refining answers with contextual reasoning. It prioritizes factual accuracy, linking directly to sources when confidence is low. Early tests show users enter twice as much detail and follow up 25% of the time.

Currently in Search Labs for Google One AI Premium subscribers, AI Mode is available via google.com/aimode and the Google app. Google is also expanding AI Overviews to cover advanced coding, math, and multimodal queries, now open to all users. The update positions Google Search against ChatGPT Search and Perplexity AI in the AI-driven search race. Read more.

'Zuchongzhi 3.0' launched: China sets new quantum computing benchmark

Chinese scientists have introduced Zuchongzhi 3.0, a 105-qubit superconducting quantum computer, pushing the limits of quantum speed. Developed by a team led by Pan Jianwei and Zhu Xiaobo, it executes quantum tasks a quadrillion times faster than top supercomputers and one million times faster than Google’s latest results.

With 182 couplers and improved qubit control, Zuchongzhi 3.0 outperforms its predecessor, reinforcing China’s role in the global quantum race. The research, published in Physical Review Letters, highlights progress in quantum error correction, entanglement, and simulation. The team is working on surface code error correction, a step toward scalable quantum computing. Read more.

• Scientists create a supersolid from light for the first time

• Meta bets on smarter AI agents, real-time voice, and a new monetization push

• Tavus introduces family of AI models to power real-time human face-to-face interaction

• Hedra launches Character-3 omnimodal AI model and Studio platform for video generation

• Anthropic expands AI Console, courts Washington, and faces technical missteps

• Andreessen Horowitz releases latest Top 100 AI apps ranking, reports major market shifts

• Mistral releases new optical character recognition (OCR) API claiming top performance globally

• Scale AI secures multi-million dollar Pentagon deal for AI-driven military planning

• All the top power players at Elon Musk’s OpenAI competitor, xAI

• Pika AI rolls out new video editing features, including region control and frame interpolation 

• When outplayed, AI models resort to cheating to win chess matches

• Google co-founder Larry Page reportedly has a new AI startup

• Christie's AI art auction reportedly exceeds expectations

• Hunyuan releases open-source code for video generation AI model

• Turing Award honors researchers behind AI breakthroughs from AlphaGo to large models

• New open-source math model Light-R1-32B surpasses equivalent DeepSeek performance with only $1000 in training costs

• Watch Unitree G1 humanoid robot kick baton mid-spin in 720° turn, master combat moves

• A quarter of startups in YC's current cohort have codebases that are almost entirely AI-generated\

📄 Babel: Open Multilingual Large Language Models Serving Over 90% of Global Speakers

📄 MultiAgentBench: Evaluating the Collaboration and Competition of LLM agents ​​

📄 ABC: Achieving Better Control of Multimodal Embeddings using VLMs  

📄 START: Self-taught Reasoner with Tools 

📄 EgoLife: Towards Egocentric Life Assistant

Thank you for reading today’s edition.

Your feedback is valuable. Respond to this email and tell us how you think we could add more value to this newsletter.

Interested in reaching smart readers like you? To become an AI Breakfast sponsor, reply to this email or DM us on 𝕏!