·
·
AI / Artificial Intelligence
Anthropic/Claude
Google / Gemini
·
LLM
Anthropic/Claude
Google/Gemini
Generative AI
Image generation
AI News Week 28 u2013 The best model for the job: Sonnet 3.5 is the new standard, Claude 3.5 Opus delayed, Google's massive week
Anthropic launched Claude 3.5 Sonnet on 30 June, offering near-Opus level performance for daily business tasks. Fable 5 returned globally on 1 July, featuring a new safety filter, and is restricted to usage credits from 7 July. Google had a massive week: NotebookLM shorts, Nano Banana 2 Lite, and Gemini in the Spark Mac app (not yet available in Switzerland). Additionally, Claude Managed Agents and Claude Tag for Slack launched, while SpaceX acquired Cursor.

1. Claude Sonnet 5 – the new daily standard
Anthropic released Claude Sonnet 5 on 30 June. It is their most agentic Sonnet model yet: it plans, uses tools like browsers and terminals, and works independently across multiple steps. Performance is close to Opus 4.8 – at a significantly lower price. Compared to its predecessor, Sonnet 4.6, it shows clear improvements in reasoning, tool use, coding and knowledge work. On an agentic coding benchmark, Sonnet 5 scores 63.2 per cent, while Opus 4.8 reaches 69.2 per cent and Sonnet 4.6 sits at 58.1 per cent.
Sonnet 5 is now the standard for Free and Pro plans. It is also available for Max, Team and Enterprise, as well as in Claude Code and via the API (claude-sonnet-5). Cyber safety filters are active by default. A new tokenizer counts roughly 1.0 to 1.35 times more tokens for the same text.
Analysis: In forums, some complain that Sonnet 5 ends up costing more than Opus because it takes longer to deliver good results. This criticism comes almost exclusively from developers with pure coding tasks. For knowledge work – research, writing, analysis, summarising – Sonnet 5 is efficient and powerful. This is exactly where the value lies for most: Opus-level quality, cheaper to run.
Rule of thumb for model choice: on the Pro plan, use Sonnet 5 as your standard and switch to Opus 4.8 for complex tasks. On Max plans, use Opus 4.8 as standard, Sonnet 5 if you hit limits often, and Fable 5 only for the hardest problems.
2. Fable 5 returns – with the brakes on
An update to the story from week 27: on 30 June, the US government lifted export controls on Fable 5 and Mythos 5. Fable 5 has been globally available since 1 July. As a reminder: Washington imposed the ban on 12 June, prompting Anthropic to suspend access for everyone.
Its return comes with two constraints. First, a new safety filter: a classifier checks every request and blocks cybersecurity and some biology tasks. It redirects sensitive or ambiguous requests to Opus 4.8 – you still get an answer, just from a different model. For normal knowledge work, this block rarely triggers in practice.
Second, access: for Pro, Max, Team and selected Enterprise plans, Fable 5 is included up to 50 per cent of the weekly limit until 7 July. After that, it is only available via usage credits – at an extra cost.
Analysis: Fable 5 remains the most powerful, but also the most expensive and slowest model. It is not worth it for simple tasks. Use it selectively for your most critical, complex challenges – Sonnet 5 or Opus 4.8 are sufficient for everyday work.
3. Google delivers a monster week
While everyone focused on Anthropic, Google rolled out several updates at once.
NotebookLM gets Shorts. Alongside the familiar video overviews, you can now generate short, vertical clips in 9:16 format of about one minute. They summarise your sources using animated images, moving text and a voiceover. Announced on 30 June, initially for paid Ultra and Pro subscriptions on web and mobile, powered by the new Nano Banana 2 Lite image model. Practical detail improvements are also here: you can select sources directly within each studio tool, explainer videos host custom styles, and NotebookLM suggests fitting topics based on your sources.
Nano Banana 2 Lite. Google released a fast, cheap variant of its image model. Quality is lower than regular Nano Banana 2 – details are softer, background faces blur – but it generates in seconds. In Gemini, select "Flash Lite" to use it. For image editing or reference images, regular Nano Banana 2 remains the better choice. The benefit of the Lite version lies in speed and cost, especially for developers embedding image generation.
Omni Flash for developers. The video model Gemini Omni Flash, available in Gemini and Flow since Google I/O, is now accessible via the Gemini API and AI Studio. Developers can integrate video generation directly into their own applications.
Gemini Mac App and Spark. The Mac app offered little for a long time. Now comes Notebooks for organising chats, similar to projects. More importantly: Spark, Gemini's cloud-based agent (Ultra only), is now in the Mac app and can connect to local folders – without requiring your computer to be running. This approaches Cowork or Codex territory, but is not there yet. New Spark integrations: Canva, Dropbox, Instacart, OpenTable, Zillow Rentals, Google Tasks, Google Keep and support for custom MCP servers. As is often the case: none of this is available in Switzerland yet. Google continues to struggle with timely international rollouts of its innovations.
Gmail Live (preview). You can chat with Gmail via voice – to sort, reply to and scan emails – similar to Gemini Live.
Analysis: Google is expanding horizontally, not in the depth of a single model. The most practical features are NotebookLM Shorts as a quick explainer format and Spark with local folder access – Google is aiming directly at agentic desktop tools.
4. Anthropic expands its enterprise models
Alongside the two model launches, Anthropic released several tools for business use.
Claude Managed Agents are available in public beta. You can build your own agents without managing the infrastructure: you define the tasks, tools and guardrails, and Anthropic runs the agent on its systems. A built-in orchestrator decides when to call which tool, manages context and recovers from errors. This allows you to go live in days instead of months. The agents run independently for hours, can launch on a schedule and remember lessons from previous runs via persistent memory. Fable 5 runs here without modification.
Advisor Strategy. You can use Fable 5 as an advisor model: a cheaper worker model calls it mid-task to review the plan and output. This elevates quality without running the entire task on the expensive model.
Claude Tag brings Claude directly into Slack. Tag @Claude in a channel and assign a task – Claude breaks it down into steps, works through them using its tools and replies in the thread with the result, leaving you to focus on other work. Claude remembers the context of the channels it belongs to. It is multiplayer: there is one Claude per channel interacting with the whole team – a shared team member rather than a private chatbot. Available in beta for Enterprise and Team clients. Additionally, Claude Science, a workbench for researchers, launched on 30 June.
Analysis: The pattern mirrors the centrally managed integrations from week 26 and Mistral's enterprise updates in week 27: providers are making their AI corporate-ready. Moving away from individual setup to managed, monitored processes.
5. Briefly noted
Cursor launches iOS app – and now belongs to SpaceX. Coding tool Cursor launched an iOS app. For context: SpaceX acquired Cursor's parent Anysphere on 16 June for $60 billion in an all-stock deal (expected to close in Q3). Following February's xAI merger, this also puts Grok under the same parent company. Cursor is no longer owned solely by its four MIT founders, but is now a SpaceX subsidiary.
Claude Design now exports video. The prototyping, document and slide tool can now output designs as video files – handy for animations and motion graphics.
Google Sheets Canvas. Introduced at Cloud Next in April, now in wider release: static spreadsheets become interactive mini-apps – dashboards, Kanban boards, calendars, gallery views – built via prompts to Gemini.
Three things to watch this week
1. Switch your standard to Sonnet 5. For knowledge work, you get Opus-level quality cheaper to run. Save Opus 4.8 for complex work and Fable 5 only for the hardest tasks – especially as Fable 5 incurs extra costs from 7 July.
2. Expect blocks and extra charges with Fable 5. The new safety filter redirects sensitive queries to Opus 4.8, and from 7 July, Fable will only run on usage credits. You do not need it for everyday tasks.
3. Test NotebookLM Shorts. A 60-second vertical video from your own sources is a fast explainer and content format – useful for both internal knowledge sharing and social media.