In partnership with
Welcome, AI enthusiasts
Elon Musk's SpaceXAI just launched Grok 4.5, and it beat Anthropic's Opus 4.8 on three of five coding benchmarks for under half the cost. That undercuts the labs that set the going rate for AI coding help. For anyone who writes code, the strongest tools on the market keep getting cheaper. Let's dive in!
In today’s insights:
Musk's Grok 4.5 Beats Opus 4.8 for Half the Cost
America Loses Nobel Prize Scientist to China's AI Lab
OpenAI Just Broke AI's Top Coding Scoreboard
Read time: 4 minutes
LATEST DEVELOPMENTS
Evolving AI: SpaceXAI's Grok 4.5 arrived today as its strongest model for coding and agents.
Key Points:
Grok 4.5 topped every model on the SWE Marathon coding test at 29%, finishing above both Opus 4.8 and Fable.
It also beat Opus 4.8 on Terminal Bench and DeepSWE 1.0, winning three of the five benchmarks SpaceXAI published.
SpaceXAI says the model runs under half the cost of comparable systems, at $2 per million input tokens and $6 per million output.
Details:
Grok 4.5 trained alongside Cursor on tens of thousands of NVIDIA GB300 chips and is now available in Grok Build and Cursor on every plan, where SpaceXAI set it as the default model and opened free access for a limited time. On SWE Bench Pro it resolves tasks with roughly a quarter of the output tokens Opus 4.8 needs. The model builds working apps from a single prompt, with EU availability expected in mid-July.
Why It Matters:
Cursor users can now run a frontier coding model at a fraction of what the leaders charge, which reshapes the daily math for freelancers and small teams who bill by the hour. Cheaper models like it mean the best coding help is no longer gated behind the biggest budgets. A student can now leave it running all day.
2 DAY LIVE AGENTIC AI WORKSHOP (BY EVOLVING AI), JULY 11-12
🚀 Build AI Agents in One Weekend
Evolving AI: Everyone has access to the same AI models. The difference is what you build around them.
If you're still copying prompts into ChatGPT, you're missing what comes next: AI agents, MCP servers, and autonomous workflows that can access live data and complete tasks for you.
In our live 2-Day Agentic AI Workshop (July 11–12), you'll build everything from scratch alongside us, no prior experience required.
What you'll build:
🤖 AI agents that automate real tasks
🔌 Your own MCP server connected to live data
⚡ Multi-agent workflows that collaborate autonomously
🛠 Skills, code, templates, and recordings you can keep using long after the workshop
By Sunday evening, you'll have working AI systems running on your own machine, not just notes from another AI course.
Source: Getty Images
Evolving AI: Anthropic landed a Nobel scientist as Google's AI talent thins.
Key Points:
Omar Yaghi, who shared the 2025 Nobel Prize in Chemistry, is now a full-time professor at Tsinghua University in Beijing.
He will direct a new institute built to use AI for discovering new materials.
The move follows steep Trump administration cuts to US research grants and fresh limits on foreign scientific collaboration.
Details:
China's Tsinghua University marked Yaghi's arrival with a 3 July ceremony, four years after he first joined as an honorary professor in 2022. His reputation rests on metal-organic frameworks, porous compounds that can pull drinking water from dry air and capture carbon, of which chemists have now built more than 100,000. He had worked at UC Berkeley since 2012 and still runs the California startups Atoco and WaHa.
Why It Matters:
Omar Yaghi's materials already pull drinking water from desert air, and he is betting AI can find the next ones faster. For readers that means quicker progress on clean water and medicine, now steered from Beijing. He built his career in American labs, and his next discoveries will come from a Chinese one.
Watch how owning AI deployment expands your career
Missed the live roundtable? Watch three teams share how they made AI ownership their job, now on-demand.
AI BENCHMARK
🧪 OpenAI Just Broke AI's Top Coding Scoreboard
Evolving AI: OpenAI audited the test used to rank AI coders and says a third of it is broken.
Key Points:
SWE-Bench Pro was OpenAI's recommended coding test, chosen after it retired an earlier one.
The audit found the test broken in several ways, with some tasks rejecting correct code and others passing incomplete fixes.
OpenAI is now retracting that recommendation and warning developers to check results carefully.
Details:
Scale AI built SWE-Bench Pro to track agentic coding across 731 public tasks, and frontier models had climbed from 23.3% to 80.3% on it in eight months. OpenAI ran a review pipeline and five experienced engineers over the same set. The pipeline flagged 27.4% of tasks as broken and the engineers 34.1%, and OpenAI settles on about 30% overall. The flaws ran from over-strict tests to vague or misleading prompts.
Why It Matters:
OpenAI's finding reaches anyone who picks an AI tool by its ranking. The scores that tell a developer which assistant to trust may rest on tasks that were never scored fairly. Those same numbers get cited as proof AI is catching up to human coders. For now, they sit on ground OpenAI no longer trusts.
👀 Click on the image you think is real
QUICK HITS
🔧 Anthropic in Early Talks With Samsung to Manufacture Its First Custom AI Chip.
🤝 Globant and Vercel Launch AI Pods to Take Enterprises From Agentic AI Pilots to Live Production.
🔬 Syntiant Files for a Nasdaq IPO, Betting Investors Want Edge AI Too.
🩺 a16z Leads $110M Into Pearl Health as Value-Based Care AI Turns Profitable.
📈 Trending AI Tools
🤖 Lindy – The simplest way for businesses to create, manage, and share agents*.
🧠 Grasppy - Turns scattered AI chats across 17 platforms into searchable project memory.
🔎 Gemmetric - Measures and improves how AI systems recognize and recommend your business.
🗂️ Datascale - AI-native data design tool for diagrams, lineage, and ER models.
*partner link