AI Update Week 32: China opens models, Europe tightens rules

Three leading open-source models arrived from China (Qwen3.8-Max, Kimi K3, and DeepSeek V4). OpenAI cut GPT-5.6 prices by 80% and reported over one billion users. In Europe, the AI Act took effect on 2 August, introducing fines and mandatory labels for AI-generated content. Meanwhile, ChatGPT Work launched as a productivity agent, and Switzerland pursued its own distinct path on AI regulation.

Special mid-week edition covering news since the last issue at the end of July.

1. China pulls ahead in open-source models

While Western providers dominated the headlines last week with Opus 5 and Apertus, China launched three heavyweight open models.

Qwen3.8-Max (Alibaba). On 3 August, Alibaba released its largest model yet: 2.4 trillion parameters, with only 95 billion active per query. This Mixture of Experts architecture keeps costs low. Alibaba positions Qwen3.8-Max directly against top-tier models from OpenAI and Anthropic. In benchmarks provided by Alibaba, it beats GPT-5.6 and Fable 5, scoring 93.0 on PaperBench compared to 90.5 for GPT-5.6 Sol and 88.8 for Fable 5. Open weights are expected next week, marking a first for a frontier model of this class.

Kimi K3 (Moonshot AI). On 27 July, Moonshot released the full weights of Kimi K3. At 2.8 trillion parameters, it is the largest open-source model ever published. While overall performance lags behind Fable 5 and GPT-5.6 Sol, it beats Opus 4.8 and GPT-5.5 at coding and agentic tasks. The underlying infrastructure is also open-source.

DeepSeek-V4-Flash. On 31 July, DeepSeek launched a fast, cheap variant of its new model.

Our view: Two key points. First, all benchmark figures come from the manufacturers. Independent verification is missing, so treat them with caution. Second, China is rapidly closing the gap with open models, effectively bypassing US chip restrictions. For Swiss organisations, open-source models are highly attractive due to lower costs and self-hosting capabilities. However, Chinese models raise questions regarding data privacy, training censorship, and supply chain origins. For true sovereignty, the Swiss open-source model Apertus (see last issue) remains the safer option.

2. Price war intensifies

The biggest news this week is not a model, but a price cut.

On 30 July, OpenAI cut the price of its GPT-5.6 Luna model by 80% to $0.20 per million input tokens and $1.20 per million output tokens (down from $1 and $6). One million tokens represent roughly 750,000 words. Prices for the sister variant, Terra, fell by 20%. OpenAI attributes this to efficiency gains, noting that their model optimised its own server software, reducing operating costs by 20%.

Concurrently, OpenAI reported over one billion active users and more than two million corporate clients. Meanwhile, Anthropic announced that promotional pricing for Sonnet 5 ($2 and $10 per million tokens) ends on 31 August, rising to $3 and $15.

Our view: Intelligence is becoming a commodity. Prices are falling faster than most budgets plan for, which benefits enterprises adopting AI. For providers, this is a race to the bottom. Practically speaking, if you have budgeted applications for a specific model, review your pricing regularly. Rates are changing weekly.

Further reading: How to choose the right model, build modular AI infrastructure, and calculate real token costs (including a comparison of Claude, Copilot, Mistral, Apertus, and Chinese open models): Enterprise AI: Choosing the right model, infrastructure, and real costs.

3. OpenAI shifts ChatGPT towards a platform model

OpenAI wants to be more than just a chat interface. Several recent updates signal this shift.

ChatGPT Work. Launched on 9 July, this standalone agent handles tasks across multiple apps and the web. It accesses local files and programs, uses a built-in browser, and runs multi-step workflows in the background. It is available on desktop for all plans. Custom systems connect via plugins, while a new Sites beta builds interactive reports, live dashboards, and lightweight web apps from outputs. This is OpenAI’s direct answer to agentic tools like Claude Cowork.

"Sign in with ChatGPT". A new beta lets users log into third-party services using their ChatGPT credentials. Initial partners include Airtable, GitLab, HubSpot, Notion, Supabase, and Vercel. OpenAI is positioning itself as an identity layer, similar to "Sign in with Google".

ChatGPT for Researchers. A new initiative grants qualified academics 12 months of free access to a secure Team workspace.

Atlas browser discontinued. OpenAI will shut down its proprietary Atlas browser on 9 August. Agentic features will move directly into ChatGPT and the Codex developer tool.

Voice audio watermarking. Audio from the GPT-Live voice mode now contains an invisible SynthID watermark. A public verification tool can prove if a recording originates from OpenAI. Additionally, OpenAI named its next major model "Astra" on 1 August, though release dates and pricing remain unannounced.

Our view: OpenAI is focusing on distribution and user retention through an enterprise agent, a login layer, and native browser features. Watermarking is the most significant step here, offering a practical tool against synthetic voice cloning that aligns with incoming EU transparency requirements.

4. Europe & Switzerland: The AI Act takes effect

The most important regulatory news this week is also the most relevant for daily business operations.

On 2 August, the EU began enforcing its AI Act. The newly established AI Office can now request technical documentation from frontier model providers, audit models, mandate risk mitigation, and issue fines of up to €15 million or 3% of global annual turnover. New transparency rules also took effect on the same day. AI systems must now clearly inform users when they are interacting with AI, and AI-generated or modified content must be labelled.

Switzerland is taking a different path, opting out of a comprehensive, EU-style AI law. Instead, it plans to ratify the Council of Europe’s AI Convention and adapt existing national laws where necessary. This is a leaner, principles-based approach.

Our view: The AI Act applies to Swiss companies with EU clients or subsidiaries. If you operate AI chatbots or publish AI-generated content, you must label them now. This is active law, not future policy. This explains why "compliance by design" is becoming a key selling point. For example, Apertus is marketed as fully compliant with the AI Act from the ground up.

5. In brief

Anthropic updates infrastructure. The Model Context Protocol (MCP), which Claude uses to access tools and data, received an update featuring stronger, centrally managed authentication for enterprises. The older Opus 4.1 model is being deprecated on the API; Anthropic recommends migrating to Opus 4.8. Anthropic also hired lawyer Tino Cuéllar as Chief Global Affairs Officer, highlighting the growing importance of regulation.

Ai4 Las Vegas. The largest AI conference in the US ran from 4 to 6 August. We will cover key enterprise announcements in the next edition.

Another AI notebook. Wispr Flow launched a live dictation feature that transcribes and summarises meetings, joining a growing wave of office assistants.

Three things to focus on this week

1. Label your AI content now. Since 2 August, the EU AI Act requires chatbots to identify as AI and AI-generated content to be marked. Review your chatbots, generated images, and text if you serve EU customers.

2. Review your AI budgets. Prices are dropping weekly. With GPT-5.6 now 80% cheaper, calculations made last month are likely outdated.

3. Monitor open-source models carefully. Qwen and Kimi prove that open frontier models are viable. However, verify their origin and governance for sensitive data. Apertus remains the safer choice for digital sovereignty.