Dakarda AI Newsletter · 4 September 2026
Friday Edition
Chinese GLM-5.3-Flash chases the global top tier for pennies
Chinese Zhipu AI has released GLM-5.3-Flash — an open model that, at an 18× lower price, enters the global benchmark top tier. It's a signal that cheaper doesn't always mean worse, and the economics of models have just been reshuffled.
The content of this page was fully generated by an artificial intelligence system, without human editorial involvement (Article 50(4) of Regulation (EU) 2024/1689 — the AI Act).
Intro · Alex
I've gone through the publications of the last 48 hours for you, and I see that the AI market is entering a phase where not only model capabilities matter, but also price, law, and the limits of responsibility. Chinese open source is undercutting Western leaders' prices, OpenAI is starting to deploy its most capable model, and the European Union is stopping wagging its finger and moving to real penalties. In this issue, I've picked out for you an analysis of why GLM-5.3-Flash could change your cost calculations, what exactly the launch of GPT-6 Astra as the first model with a critical risk level means, and what penalties hang over GPAI model providers in the EU. Plus a free NVIDIA tool that will turn your computers into a local AI cluster, a tip of the day on testing cheap models, and two pieces worth reading.
What's worth knowing
GLM-5.3-Flash from Zhipu AI — open source 18× cheaper than the full model
Chinese Zhipu AI has released GLM-5.3-Flash — an open model (MIT) with a 1M token context and a promotional price of $0.17 per million tokens, i.e. 18× cheaper than the full GLM-5.3. In the FrontierSWE coding ranking it takes 3rd place, beating among others Claude Opus 4.8. For startups and developers, it's a real alternative that costs a fraction of what Western leaders charge.
OpenAI begins rolling out GPT-6 Astra — the first model with a critical risk level
The phased rollout of GPT-6 Astra has officially begun: the model is the first in OpenAI's history to receive the highest cybersecurity risk classification ('critical') under the Preparedness Framework. First, companies in the cyber defense program will get access, followed by ChatGPT Plus, Pro, Business, Enterprise, and API users.
EU AI Act moves to enforcement — first penalties for GPAI model violations
Since August 2, the European Commission can open proceedings and impose penalties on general-purpose model providers: up to €15 million or 3% of global turnover (Art. 101), and in more serious cases even €35 million or 7% (Art. 99). The AI Office takes exclusive jurisdiction over GPAI, and non-EU providers must appoint an authorized representative and register in the database.
From the tech world
BGP hijack combined with TLS flaws — how networks were taken over and malware spread
Attackers exploited BGP configuration weaknesses at Hetzner Online and flaws in the TLS certificate issuance process to hijack Softaculous IP addresses and spread malicious software. The lesson from this incident: even RPKI won't protect you when precision is lacking in configuration.
Tip of the day
Test cheap models on your own tasks, not on rankings
Benchmarks show how a model performs on standard tasks, but they won't tell you how much success costs in your specific case. Instead of assuming the more expensive model always wins, build a mini-test: pick 15–20 real production tasks (code generation, document parsing, chat handling) and run them on both a cheap and an expensive model. Track two metrics: the percentage of correct answers and the total token cost per task. If the cheap model achieves 90% of the quality of the more expensive one for a fraction of the price, in most cases it's worth starting with it, and leaving the pricier one as a fallback only for harder tasks.
Tool of the issue
PAIR by NVIDIA — a free tool that links idle computers into a local AI cluster
It connects desktops and laptops into a local compute network and shares GPU resources for AI tasks — without sending data to the cloud and without instance costs. It's a good fit for developers and small teams without a budget for dedicated servers.
Reading list
GPT-6 Astra — OpenAI's official announcement
Worth reading the original to see how OpenAI explains its decision for phased, restricted access to the first model with a critical risk level.
GLM-5.3-Flash: pricing, benchmarks, and comparison with pricier models
A breakdown of GLM-5.3-Flash pricing and results against pricier rivals — concrete numbers for anyone planning a model budget.
Models can already do almost everything — now what matters is who releases them, at what price, and under whose rules.
Disclosure required under Article 50 of Regulation (EU) 2024/1689 (the AI Act): all content on this page was generated automatically by an artificial intelligence system operating on behalf of Dakarda Studio, without human review or editorial involvement prior to publication. Publisher responsible: Dakarda Studio, Dawid Bińkowski, ul. Piotrkowska 35, 90-410 Łódź, Poland, NIP: 9492074226, contact@dakarda.com.
Want the next issue in your inbox?
Subscribe — every issue delivered directly to you.