AI News of the Week (Week 29/2026): GPT-5.6 from $1, Gemini Delayed, Anthropic Builds Its Own Chip

Every week, more happens in AI than any business owner can casually keep up with. That is why every Monday we break down the past week's most important news: soberly, in plain language, with a clear answer to the question that matters. What actually affects your business, and what is just a headline? Each item gets a relevance rating from 1 to 5 for SMEs (small and medium-sized enterprises).

OpenAI releases GPT-5.6 in three sizes, from $1

OpenAI has unveiled GPT-5.6, for the first time in three sizes: Luna, Terra, and Sol. All three models offer a 1 million token context window, and the entry model starts at $1 per million tokens.

What it means: The price war is on. For businesses, this is the week's most important news, because falling token prices mean automation use cases that were too expensive six months ago are worth revisiting now. Recalculate your use cases. Relevance: 5 out of 5.

Google delays Gemini 3.5 Pro by months

The planned frontier launch has slipped significantly. Internal testing revealed weaknesses in coding and complex reasoning.

What it means: Announced is not the same as delivered. The real race happens between demo and daily use, and anyone who builds their AI plans on roadmap promises is building on sand. Decide based on what actually works in your use case today. Relevance: 2 out of 5.

Anthropic builds its own AI chip with Samsung

Anthropic is in talks with Samsung over an AI chip custom-built for Claude. In parallel, preparations for an October IPO are underway.

What it means: Providers are gearing up for permanent, large-scale operation. Custom chips mean designs built for availability and falling costs, not for demos. That is good news, medium term, for anyone putting AI to productive use. Relevance: 2 out of 5.

OpenAI ahead of its IPO

A confidential filing is expected in the coming weeks, with a valuation of roughly $730 billion.

What it means: A valuation is not a feature. For the product you actually use, an IPO changes nothing. Relevance: 1 out of 5.

Open model Kimi K3 beats GPT-5.6 on code

The open model Kimi K3 from Moonshot AI wins a frontend coding ranking with a 76 percent win rate, ahead of both Claude and GPT-5.6.

What it means: Open source is moving into the office. Open models run on your own infrastructure, so data stays in house. For companies with strict data protection requirements, this is the most interesting development of the week. Relevance: 4 out of 5.

Google caps Meta's compute

Meta wanted to buy more Gemini capacity than Google can supply. Google is limiting access.

What it means: It is not the models that are scarce, it is the compute behind them. If even giants get rationed, the lesson for SMEs is even clearer: build around your use case, not around a single vendor. Relevance: 3 out of 5.

What to take away

This week shows two opposing trends: usage is getting cheaper (GPT-5.6, open source), while the infrastructure behind it is getting scarcer (chips, compute). Both point toward the same course of action: calculate and implement concrete use cases now, while planning independently of any single vendor. For what good guidance on this looks like, read our guide on AI consulting for SMEs.

This briefing runs every Monday, also available in a more compact carousel format on Instagram. If you want to know which of these developments could concretely save your business time or money: talk to us.