MODELS

China's newest AI model costs 87% less than DeepSeek

DeepSeek rattled global markets in January by demonstrating that China could build competitive AI on a budget. Now, Beijing startup Z.ai is making DeepSeek look expensive.

The company's new GLM-4.5 model costs just 28 cents per million output tokens compared to DeepSeek's $2.19. That's an 87% discount on the part that actually matters when you're having long conversations with AI. We recently discussed how the further along in the conversation you are, the more impact it has on the environment, making this topic especially interesting.

Z.ai CEO Zhang Peng announced the pricing Monday at Shanghai's World AI Conference, positioning GLM-4.5 as both cheaper and more efficient than its domestic rival. The model runs on just eight Nvidia H20 chips (half what DeepSeek requires) and operates under an "agentic" framework that breaks complex tasks into manageable steps.

This matters because Zhang's company operates under US sanctions. Z.ai, formerly known as Zhipu AI, was added to the Entity List in January for allegedly supporting China's military modernization. The timing feels deliberate: just months after being blacklisted, the company is proving it can still innovate and undercut competitors.

The technical approach differs from traditional models, which attempt to process everything simultaneously. GLM-4.5's methodology mirrors human problem-solving by outlining the steps first, researching each section and then executing.

Performance benchmarks suggest this approach works:

• GLM-4.5 ranks third overall across 12 AI benchmarks, matching Claude 4 Sonnet on agent tasks

• Outperforms Claude-4-Opus on web browsing challenges

• Achieves 64.2% success on SWE-bench coding tasks compared to GPT-4.1's 48.6%

• Records a 90.6% tool-calling success rate, beating Claude-4-Sonnet's 89.5%

The model contains a total of 355 billion parameters, but activates only 32 billion for any given task. This reliability comes with a trade-off: GLM-4.5 uses more tokens per interaction than cheaper alternatives, essentially "spending" tokens to "buy" consistency.

Z.ai has raised over $1.5 billion from Alibaba, Tencent and Chinese government funds. The company represents one of China's "AI Tigers," considered Beijing's best hope for competing with US tech giants.

Since DeepSeek's breakthrough, Chinese companies have flooded the market with 1,509 large language models as of July, often using open-source strategies to undercut Western competitors. Each release pushes prices lower while maintaining competitive performance.

This feels like China's tech strategy crystallizing into something Washington should worry about. Z.ai just proved that a sanctioned company can deliver state-of-the-art AI while undercutting everyone on price. The 87% cost reduction targeting output tokens specifically shows an understanding of where users feel pricing pain most.

Most concerning for US policymakers is that export controls clearly aren't slowing Chinese AI development. A blacklisted company just delivered competitive performance while operating under restrictions designed to cripple such capabilities.

TOGETHER WITH MISO

From a CalTech Garage to a $1T Market

In 2016, some innovative CalTech robotics students set out to transform fast food. That first garage-built prototype became Miso Robotics.

Since then, Miso’s robots have worked 200K+ hours in live kitchens for brands like White Castle, frying 4M+ baskets of fries, chicken, and more. That traction convinced NVIDIA and Amazon to help refine Miso’s AI and robots. As restaurants grapple with 144% turnover rates and $20/hour minimum wages, that’s a major edge. 

No wonder Miso’s first fully commercial robot, Flippy Fry Station, sold out initial units in one week. 

Now, Miso’s going nationwide and scaling its US-based manufacturing across a $4B/year revenue opportunity. And you can share in its growth as an investor. But you only have until August 14 to secure the current share price.

Get up to 8% bonus shares with your Miso investment before Miso’s share price changes

APPLICATIONS

AI reaches gold level at math olympiad, but 26 humans still beat them

Both OpenAI and Google DeepMind achieved gold medal-level performance at the 2025 International Mathematical Olympiad, but 26 human students scored higher than the AI systems, with five achieving perfect 42-point scores.

The AI models each scored 35 out of 42 points on the world's most prestigious high school mathematics competition, placing them around 27th among 630 global contestants. Google's advanced Gemini model with Deep Think solved five of six problems perfectly, with results officially verified by IMO organizers.

OpenAI's experimental reasoning model also scored 35 points, but didn't officially enter the competition. Researcher Alexander Wei announced the results on social media before verification, drawing criticism from Google researchers who accused OpenAI of premature announcement.

• The IMO features six exceptionally difficult problems across algebra, combinatorics, geometry and number theory

• Only about 8% of human contestants typically earn gold medals in the annual competition

• Both systems solved problems requiring sustained creative thinking over hours

Many observers believe this may be the last time humans outperform AI at the prestigious competition. The rapid progression from Google's silver medal just one year ago suggests AI mathematical capabilities are advancing faster than anticipated, marking a potential inflection point in machine reasoning abilities.

TOGETHER WITH WRITER

A new kind of agent for enterprise work

Complete projects in minutes rather than weeks — Action Agent coordinates and executes complex work for you. 

It’s a researcher working around the clock, a strategist devising go-to-market plans, a data scientist providing in-depth analyses, and a programmer building and deploying websites. Just give it a goal and Action Agent puts together a multi-phase plan, then automatically executes to deliver exactly what you need: in-depth reports, polished slide decks, detailed spreadsheets, and more. 

Try the super agent for enterprise work:

• Creates real deliverables, not first drafts  

• Connects with any tools, data, and systems  

• Keeps working even when you’re offline  

• Fully transparent and self-correcting

Give Action Agent a project to tackle today.

Try the beta for free

POLICY

Cadence pays $140 million for illegal chip tool sales to Chinese military

Silicon Valley chipmaker Cadence Design Systems agreed to plead guilty and pay more than $140 million to resolve charges that it illegally sold electronic design automation tools to a Chinese military university involved in nuclear weapons simulation, the Justice Department announced Monday.

Cadence violated export controls by selling chip design software and hardware to front companies representing China's National University of Defense Technology between February 2015 and April 2021. NUDT had been on the Commerce Department's Entity List since 2015 due to its use of U.S. technology to build supercomputers supporting nuclear explosive simulation.

• Cadence exported EDA tools at least 59 times to Central South CAD Center, an alias for NUDT, despite knowing the university's restricted status

• Company employees installed hardware directly on NUDT's campus and provided software downloads through Cadence portals

• The violations continued even after Cadence learned CSCC was an alias for the restricted university

The investigation began over four years ago when Cadence received subpoenas from the Commerce Department in February 2021, followed by Justice Department subpoenas in November 2023.

China represented 12% of Cadence's revenue last year, down from 17% in 2023. The company will face three years of probation as part of the plea agreement.

LINKS

• Anthropic unveils new rate limits to curb Claude Code power users

• Tesla, Samsung sign $16.5b deal to make AI chips

• The surprise winner in the U.S.-China tech war? Vietnam’s chip industry

• AI cloud startup Fireworks discusses $4b valuation

• Nvidia-backed Cohere forms AI alliance with telecom firm BCE

• Seriously, why do some AI chatbot subscriptions cost more than $200?

• ‘The Wizard of Oz’ blown up by AI for giant Sphere screen

• Graphed: Upload data sources, one-shot generate dashboards, and chat with your AI data scientist to get insights

• Trickle: Turn your ideas into live apps and websites with AI.

• Chatbase: Build and deploy AI support agents on 1 platform

• AgentOps: Developer platform for building AI agents and LLM apps

• Synthesia: Lead Enterprise Product Marketing Manager

• UiPath: Transformation Analyst

• Captions: Software Engineer, Backend

• Mistral AI: Account Executive - AI for Citizens

GAMES

A QUICK POLL BEFORE YOU GO