GLM-5.3 democratizes cybersecurity, but its developer can’t guarantee against misuse. (Picture: Shutterstock)The Chinese open weights model surpasses GPT-5.6 Sol and Fable 5, the sanitized Mythos model, on CyberGym — a widely watched benchmark for cybersecurity — as Z.ai says defensive cyber functions should not be «the privilege of the few,» according to The South China Morning Post.
GLM-5.3 scored 84.5% on the benchmark, slightly higher than Fable 5’s 83.8% and GPT-5.6 Sol’s 83.6%.
On ExploitBench, which is more a measure of a model’s offensive ability, the model trailed the frontiers with 54.5% versus 78% for Fable 5 and 76.5% for GPT Sol.
Z.ai, the lab behind the model, says they will release it as open source with open weights in about two weeks, according to Reuters, after first letting «selected security partners» evaluate it in «controlled settings,» Z.ai writes on security.
The new V4 Pro is one of the best open weights models, and comes with a low price. (Picture: generated)The stealthy upgrade to version 0831 yesterday had many excited, and there was an impressive screenshot of benchmarks circulating and creating buzz all day long, but nobody has been able to verify its origin.
What is confirmed is that DeepSeek has indeed updated its V4 Pro model for the first time since April, and just recently posted this message to its website:
«🎉 The official version of DeepSeek-V4-Pro has been released, featuring significantly enhanced agent capabilities and support for the Responses API and Codex integration. It is now fully available across the web, mobile app, and API; we welcome your testing and feedback.»
DeepSeek V4 Flash is currently the cheapest GPT-5.5-level offering out there. (Picture: Shutterstock)UPDATE: DeepSeek has now increased their API prices with a system of peak and off-peak prices for both the Pro and Flash models. It’s up to $1.32/M in and $3.96M/out for peak Pro usage, going into effect on August 17. You can see the new prices here.
The Chinese AI lab recently launched DeepSeek V4 Flash, a model competitive against Gemini 3.5 Flash and GPT-5.5 at a fraction of the cost, but that may be coming to an end.
According to Reddit user AlyoshaV and later confirmed by Bloomberg, DeepSeek sent a notification to its users on Thursday notifying them of their intention to «raise the overall pricing for DeepSeek API services in the near future.»
They are not saying how much they intend to raise prices from the current $0.14/$0.28 for a million tokens in/out, but they do say the increase will be «significant.»
Many were wondering if DeepSeek could keep up this pricing structure amid rising popularity in an increasingly cost-conscious market, due to their lack of significant compute power.
DeepSeek hasn’t publicly announced just how much compute they currently have online, believed to be a mixture of Huawei and Nvidia H800 chips, but they are currently investing in a new 1-gigawatt data center in Ulanqab in Inner Mongolia at an approximate cost of $50 billion on the free market.
Even a small price increase would still put V4 Flash on the cheaper end of the market, with the nearest competitor being GPT-5.6 Luna at double the current price.
The model lands precisely one point behind Luna on the Artificial Analysis benchmark, and is a «a significant step up from the previous generation,» the DeepSeek V4 Flash (40), AA writes.
It has a one million token context window, and has jumped from forty to fifty points on the AA evaluation and is now on par with Gemini 3.6 Flash, just behind GLM-5.2, Muse Spark, and GPT Luna.
The biggest selling point is the price, which is at $0.14/$0.28 for one million tokens in and out, far below anything on the market for this kind of performance/cost.
Compared with Luna’s newly reduced prices of $0.30/$1.20, it ticks in at about half the cost for input tokens and around a fourth of the cost for output.
Chinese open source models are reaching frontier capabilities, worrying the US administration. (Picture: generated)As the newly released breakthrough models Qwen 3.8, Kimi 3 and the upcoming DeepSeek v4 enjoy a moment in the sun, the USA is considering its options.
The Kimi 3 model is already running into compute problems due to popular demand and has decided to halt new subscriptions, as users are turning to cheaper open source models that are on par with the leading American closed source frontiers.
According to Axios, action being considered by the White House includes an executive order holding US companies liable for the risks of running open source models, putting out an advisory against them, or simply adding them to an «Entity List» that requires a license for their use.
Sources in the administration are primarily concerned about increasing capabilities in cybersecurity, but also have fears of backdoors and a general lack of security with open source models, Axios writes.
This comes hot on the heels of OpenAI’s Dean Ball’s tirade against open source on X.com, where he exclaimed his surprise that China would take the risk of a free-for-all in models this capable, as reported by Gizmodo — and called open source models decelerationist, signaling a general opposition to them from the frontier labs.
Alibaba have been accused of massively copying Claude’s responses. (Picture: Alibaba, generated)The Chinese onslaught continues, just days after the launch of groundbreaking Kimi 3, with Alibaba claiming that their 2.4 trillion parameter open source model matches the American frontier and is only behind Anthropic’s Fable 5.
There is very little information out on the model save for Alibaba’s X post, where they claim the model is «continuously evolving» and is «one of the most powerful model[s] available today.»
There are no benchmarks to back that claim just yet, and the last Alibaba model on Arena.ai’s leaderboard is the Qwen 3.7, hovering around 17th place on coding and 18th on agentic tasks. The new Qwen 3.8 would significantly improve on that performance.
GPT-5.6 was released in late July, while Fable 5 was announced in early June, putting the current window to cutting edge Chinese open models at a little over one month. And we are still waiting on DeepSeek v4, which by early leaks seems to be frontier-level, too.
Qwen 3.8 «preview» is already out and ready to test on Alibaba’s token plan, and they are promising to release the open weights «soon,» writes The Decoder.
Like Moonshot AI, behind the Kimi 3 model, Alibaba was accused of a massive distillation campaign on Anthropic’s Claude, copying some 28.8 million exchanges between April and June this year.
With Kimi 3, Chinese models are rapidly advancing toward the frontier. (Picture: Moonshot AI)The window between the American frontier models and Chinese open models keeps shrinking, and with today’s launch, Moonshot AI’s Kimi K3 is right at the edge.
In their published benchmarks, the model not only beats GPT-5.5 and Opus 4.8 in most tests, but sometimes goes right up to Fable and GPT-5.6 capabilities — at times even beating them outright, as on Arena.ai’s Code Arena.
It also drops eyebrow-raising scores in self-published benchmarks in coding, general agent use, knowledge work and visual tasks, mostly just right below the top tiers.
The insanely popular custom chatbots become a thing of the past in China. (Picture: generated)The popular chatbots Doubao and Qwen are shutting down their personalization features after China enacts the world’s first regulation of «anthropomorphic» AI agents.
They had both offered the ability to create custom agents and chatbots that would establish emotional bonds with users, and act as humanlike companions.
As of July 15, this becomes illegal in China, after the awkwardly specifically named «Interim Measures for the Administration of Artificial Intelligence Anthropomorphic Interaction Services» goes into effect.
This is the first regulation of its kind, and bans bots that «simulate human personality traits, thinking patterns and communication styles to provide sustained emotional interaction,»The South China Morning Post writes.
The personalized service of the two chatbots were used by millions, many of them kids, and often replaced real-life human interaction, as one in seven young adults in China used them to form romantic relationships, Decrypt.io writes.
UNICEF has hailed the new law, calling it a «pioneering policy [that] marks a significant global step towards regulating AI powered emotional interaction services.»
China is increasingly worried about technology leaks and talent poaching. (Picture: Adobe)Sources are telling Bloomberg that China has established a list of AI engineers that will now have to ask permission from authorities before traveling abroad.
The move comes after seeing the astronomical wages being offered for valuable AI competence by US labs, such as Meta, Bloomberg writes, but notes that China sees AI labs as a strategic asset and are concerned about data leaks.
The previous policy included restrictions for individuals that were senior researchers in education, nuclear scientists and top executives of government companies, Tom’s Hardware writes.
Before this, AI workers on the list only had to report where they were traveling, but did not have to ask for specific permission.
The new rules might inspire early-stage talent with international ambitions to leave the country before getting added to the list, and dissuade overseas talent from moving home to China, Bloomberg says.
Markets have become accustomed to roaring earnings beats from Nvidia. (Picture: Nvidia)Markets were lackluster on the last quarterly report of $68.13 billion in revenue for the AI chipmaker, as revenue growth seems to be slipping, Reuters reports.
The full year revenue hit $215.9 billion, up 65% year-on-year, with Data Center revenue hitting a record of $62.3 billion — which is responsible for their AI chips.
Nvidia also raised its guidance for Q1 2026, and is certainly not seeing any slowdown:
Chinese attacks risk bypassing the safeguards Anthropic builds into its models. (Picture: Anthropic)Anthropic claims to have discovered industrial scale extraction of Claude data from DeepSeek, Moonshot AI and MiniMax.
The massive attacks were used to improve their own models with agentic reasoning, tool use, and coding capabilities, violating Anthropic’s Terms of Service and creating a national security risk, they say.
Distillation works by sending millions of prompts to an AI to incorporate its techniques and capabilities into their own models, drastically reducing training time and costs.
They also circumvent Anthropic’s protections for use in developing bioweapons and malicious cyber activities, Anthropic says. Once these models are open sourced, this becomes available to anyone.
— These campaigns are growing in intensity and sophistication. The window to act is narrow, and the threat extends beyond any single company or region, Anthropic writes.
Kanye West can now sing perfectly in Mandarin, thanks to ByteDance’s new video model. (Picture: screenshot)The new video generator, released on Thursday, is already being hyped by state media as bigger than the launch of DeepSeek, Reuters reports.
ByteDance has yet to publish any real numbers, but videos of Kim Kardashian and Kanye West singing in Mandarin have gone viral on Weibo and x.com with millions of views.
Nvidia chips are available in China, but users need permission to buy them. (Picture: Adobe)Not much is known about the AI inference chips, or how they compare to Nvidia’s offerings, but ByteDance is going to be making about 100,000 of them «this year,» and then scale up to 350,000 units, according to Reuters.
ByteDance has been known to work with US chip producer Broadcom, and started seriously hiring chip specialists in 2022.
The new chips are set to be produced with Samsung in a deal that includes memory chips, which definitely sweetens the deal.
Production is advanced enough that Reuters’ sources say engineering samples are due by late March, which is the last stage before production.
A spokesperson for the company does not deny the report outright, but claims the information is «inaccurate,» Reuters writes.
Most US frontier labs are developing their own chips, as is Alibaba and Baidu.
China is great at playing catch-up, but can they innovate? That’s the next challenge, Hassabis, says. (Picture: Wikipedia, CC BY-SA 4.0The CEO of Google’s DeepMind has some choice words for China in a recent podcast.
They might be «just a matter of months» begin Western capabilities, he tells CNBC.
— The question is, can they innovate something new beyond the frontier? So I think they’ve shown they can catch up … and be very close to the frontier … But can they actually innovate something new, like a new transformer … that gets beyond the frontier? I don’t think that’s been shown yet, he tells the new podcast The Tech Download.
The key tech to unlock Chinese AI is access to chips, he says, where the USA is far ahead. The US recently okay’ed exports of the powerful H200 chip from Nvidia, but reception in China has been lukewarm from authorities.
— To invent something is about 100 times harder than it is to copy it, says Hassabis on the podcast. — That’s the next frontier really, and I haven’t seen evidence of that yet, but it’s very difficult.
The H200 was getting popular in China, being miles ahead on performance. (Picture: generated)Several sources are telling Reuters that the H200 chips are not permitted to enter, and authorities have told technology execs explicitly to not purchase the chips.
The H200 was cleared by Commerce for export to China in December and got finally approved this week.
