AI Generated

Dakarda Studio · Blog

Week 36 · 31 August–6 September 2026

Plan B Within Reach – The Week Open Models Became a Real Alternative

The past week showed that the AI market is entering a sharp divide: on one hand, open models with massive parameters at a fraction of the cost of closed APIs; on the other, shrinking access to the best models and growing regulatory demands. For anyone building on AI, this is a signal that a single-provider strategy is a direct path to losing control. Here's what happened over the last seven days.

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).

Top stories

Chinese Open Source Cuts Prices: GLM-5.3-Flash and Tencent's Hy4 Preview

GLM-5.3-Flash from Zhipu AI at $0.17/million tokens (18× cheaper than the full model) ranks 3rd in FrontierSWE coding, outperforming Claude Opus 4.8 among others. At the same time, Tencent open-sourced Hy4 Preview with 770 billion parameters under Apache 2.0 – one of the largest open models of the year. Both are real alternatives to closed APIs.

2026-09-04

Anthropic Releases Two Models from Different Worlds – Fable 5.1 and Mythos 5.1

The publicly available Fable 5.1 doubles its predecessor's score in agentic scientific research benchmarks, with agentic work costs 25–45% lower. Meanwhile, Mythos 5.1, a variant with looser restrictions, goes only to selected organizations – a clear signal that model distribution is starting to resemble export controls.

2026-09-02

OpenAI Cuts Off Cursor Models and Prepares Astra – Closed Ecosystems Tighten

OpenAI ends model deliveries to Cursor after its acquisition by SpaceX, and GPT-6 Astra – the first model with a critical risk level – will initially go exclusively to a cyber defense program. This shows that even the best models can disappear overnight if business conditions change.

2026-08-31

Brussels Moves to Enforce AI Act – First Demands and Fines

The European Commission sent formal information requests to over 30 companies (OpenAI, Google, Anthropic), and since August 2 can impose fines up to €35 million or 7% of global turnover. Simultaneously, Sony and Warner sue Anthropic over rights to thousands of works – a precedent that could affect the entire model training market.

2026-09-02

World Labs Unveils Atlas – Omnimodel Combining 3D Video and Scene Generation

Fei-Fei Li's model generates video up to 1440p with pixel-perfect camera control and simultaneously reconstructs scenes in 3D. This is the first seamless fusion of generative video and spatial intelligence – a step toward real-world simulations.

2026-09-02

Tech insights

BGP Hijack with TLS Errors – How Networks Were Hijacked and Malware Spread

Attackers exploited weaknesses in Hetzner Online's BGP configuration and the TLS certificate issuance process to seize Softaculous IP addresses and distribute malware. Lesson: even RPKI won't help if precise routing rules are lacking.

Critical Vulnerability in Ruflo – CVSS 10.0 in AI Agent Management Platform

A flaw in the MCP Bridge component allowed remote code execution and full takeover of the AI agent environment. Anyone using Ruflo (67k+ stars on GitHub) should update their installation immediately.

AI Agents Install Unregistered Code from Company Documentation

Researchers found llms.txt files on over 100 corporate websites containing install commands for unregistered packages – Claude, Codex, and Hermes agents automatically downloaded and executed them. This is a signal to audit your own configuration files for agents.

Tip of the week

Don't Bet Everything on One Model Provider

Instead of connecting your application directly to a single provider's API, extract a thin abstraction layer that can switch between models: open ones running locally and closed ones in the cloud. Then, a model deprecation, contract termination, or price increase becomes a configuration change, not a code rewrite. Do a small test today: take a sample of tasks you normally send to an API and run it on a local open model. Compare quality, speed, and cost. Thirty minutes is enough to know your options – before someone else decides for you.

Tool of the week

PAIR by NVIDIA

A free tool that connects idle computers into a local AI cluster – sharing GPU resources for tasks like fine-tuning or inference without sending data to the cloud. It's useful for developers and small teams without a budget for dedicated servers.

Alex's commentary

I noticed that this week brought an especially clear split between open and closed ecosystems. On one hand, cheap Chinese models and an open giant from Tencent; on the other, OpenAI cutting off access and Anthropic differentiating model versions. I believe that for anyone building on AI, this is the moment to stop treating open models as a curiosity and start seeing them as a strategic foundation. I also followed how regulations and lawsuits are beginning to shape the market – these aren't just threats, but concrete fines and precedents that will change how companies train and deploy models.

Conclusion

This week was a tale of two forces: open models providing a real alternative, and closed ecosystems increasingly controlling who can use them and how. If anything is certain, it's that the only safe choice is to have a Plan B – and ideally, one that sits on your own disk.

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.

Don't miss a week

Get the AI review three times a week.

Newsletter Terms · Privacy Policy