New: the Lumen ebook, the history and tools of AI in one bookDiscover it →
Lumen Radio I build websites@lumenaitools01
LumenAI
IT

News

Less noise, more signal.

I pick a few stories and explain why they matter. Every item has its source; I separate official announcements from rumours (often via TestingCatalog), which deserve caution.

Live

Latest from the AI world

Updated automatically every hour from TestingCatalog, OpenAI, Google, The Verge and news on Claude, Grok and Meta AI, with AI-translated titles and summaries.

Newsletter · every Monday · free

The AI news that matters, once a week.

Picked and explained by me, with one practical tip to try right away. No spam, unsubscribe in one click.

Curated and commented

The big picture, in four lines

  1. Three families compete for the top: Claude (Opus and Sonnet 5.5), GPT-6 (Sol and Astra) and Gemini 4 Argon. Meta is back with Muse Wonder.
  2. Prices fall fast: GPT-6.1 Sol completes a benchmark task for $0.72 versus almost $6 for Opus 5.5, six points apart.
  3. Assistants are becoming software: interactive answers (ChatGPT), documents and dashboards (Claude), always-on agents.
  4. In Europe, transparency is mandatory since 2 August 2026; high-risk rules slip to 2027-2028.
  1. 15 OctData
    Arena

    Gemini 4 Argon tops the Arena text leaderboard

    Google's new model leads Text Arena with 1,524 points, ahead of Meta's Muse Wonder and Claude Opus 5.5. The result is marked preliminary: few votes so far.

    My take

    Human-vote rankings reward answer style as much as substance. I would wait a few weeks of votes before switching tools.

    Sources: Arena · Text ↗
  2. 15 OctOfficial
    Krea

    Krea 3: 2K image edits in under 8 seconds

    Krea claims 2K edits in under 8 seconds and lets you train the model on a product, person or style. From $0.03 per image.

    My take

    For e-commerce and social, a "model trained on your product" matters more than a leaderboard point.

    Sources: TestingCatalog ↗
  3. 12 OctOfficial
    Meta

    Meta launches Muse Wonder and Wonder Full for agents

    Muse Wonder is available via Meta Model API and Muse Code; Wonder Full is for Muse Code subscribers. On Artificial Analysis Wonder Full scores 56, close to Claude.

    My take

    Meta is back in the leading pack after a hard year. More competition at the top means falling prices for everyone.

    Sources: TestingCatalog ↗
  4. 12 OctRumour
    Anthropic

    Anthropic reportedly testing the successor to Opus 5.5

    Some users say Opus 5.5 requests land on a new, more capable model. Whether it becomes Opus 5.6 or Fable 5.5 is unclear.

    My take

    A rumour, not an announcement. The pace is now a jump every 2-3 months: no need to chase every version.

    Sources: TestingCatalog ↗
  5. 09 OctOfficial
    Microsoft

    Microsoft Decision-1 in Foundry for structured decisions

    A model built for routing, classification and agent controls, now available in Foundry.

    My take

    Part of a trend: small specialised models acting as traffic controllers for big ones.

    Sources: TestingCatalog ↗
  6. 08 OctOfficial
    Anthropic

    Claude: Dashboards and Motion in beta; Docs, Slides and Design for all

    Dashboards enters beta on paid plans and Motion on Team and Enterprise. Docs, Slides and Design leave beta, including on Free.

    My take

    Assistants are becoming office suites. For many small firms the free plan already covers documents and decks.

    Sources: TestingCatalog ↗
  7. 07 OctOfficial
    OpenAI

    GPT-6 reaches ChatGPT with "Intelligent UI"

    Paid plans use GPT-6 Sol, Free and Go use GPT-6 Luna. Answers can include charts, forms, maps and small interactive tools generated on the fly.

    My take

    The real change is the interface, not the model: the answer becomes a small piece of software. Great for learning; verify the maths.

    Sources: TestingCatalog ↗
  8. 07 OctOfficial
    Anthropic

    Claude Haiku 5.5: lower API prices and monthly credits

    Haiku 5.5 costs $0.10 / $0.50 per million tokens. Anthropic also cut Sonnet 5.5 cache costs and added monthly API credits to Max and Team.

    My take

    For business chatbots and high-volume automations the cost per answer is now a fraction of a cent.

    Sources: TestingCatalog ↗
  9. 23 SeptOfficial
    OpenAI

    DevDay: new plans to build and host AI apps on OpenAI

    Free, Prototype and Accelerate plans to build and publish apps directly on the OpenAI platform.

    My take

    Labs want to become the hosting layer too. Convenient, but beware of lock-in.

    Sources: TestingCatalog ↗
  10. 02 AugOfficial
    Unione europea

    AI Act: transparency duties (Art. 50) now apply

    Since 2 August 2026 chatbots and synthetic content must be disclosed. Regulation (EU) 2026/1744 postponed high-risk duties to December 2027 and August 2028, and content marking to 2 December 2026.

    My take

    The delay covers high-risk systems, not transparency. If your site has a chatbot, it must say it is an AI — like this one.

    Sources: Addleshaw Goddard ↗

TestingCatalog

Want news in real time?

TestingCatalog tracks features in testing at OpenAI, Anthropic, Google and Meta daily. Remember: a rumour is not an announcement.

testingcatalog.com ↗

You can also ask Lumen: "what came out this week?". The chatbot reads the news in real time and cites the articles.