AI News Week 19 – Claude in Photoshop, GPT-5.5 catches up, Apertus 1.5 is just around the corner

1. Anthropic connects Claude with Adobe, Blender, Autodesk and six other creative tools

Anthropic launched Claude for Creative Work on 28 April 2026. Nine new connectors link Claude directly to Adobe Creative Cloud, Blender, Autodesk Fusion, Ableton, SketchUp, Affinity by Canva, Splice, plus Resolume Arena and Wire. Claude is now positioned as the control layer for design, 3D, music and live visuals.

The Adobe connector accesses more than 50 tools in Creative Cloud, including Photoshop, Premiere and Express. The Fusion connector creates 3D models from natural-language descriptions. The Blender connector is based on MCP, so other LLMs can use it too. Anthropic also works with the Rhode Island School of Design, Ringling College and Goldsmiths, London, to provide student access.

This move should be seen in light of Claude Opus 4.7 (16 April) and Claude Design (17 April), released two weeks earlier. Anthropic is turning Claude from a chat assistant into a tool hub.

Recommendation: If you work in marketing or design with Adobe, test the connector. Today, the realistic gain is faster repetitive work, not the replacement of design work.

Sources: Anthropic: Claude for Creative Work | 9to5Mac | Neowin

2. OpenAI releases GPT-5.5 — and catches up in agentic work

OpenAI launched GPT-5.5 on 23 April 2026. The model plans multi-step tasks, uses tools, checks its own results and works through ambiguity. It was about time. Anthropic has shown this with Claude since mid-2024. We will see how much ChatGPT has caught up in practice. The API has been available since 24 April, with a 1 million token context window (400K in Codex).

The published benchmarks (all from OpenAI):

• 84.9 % on GDPval (knowledge work across 44 professions)

• 78.7 % on OSWorld-Verified (computer operation)

• 98.0 % on Tau2-bench Telecom (customer service workflows without prompt tuning)

• 82.7 % on Terminal-Bench 2.0 (Claude Opus 4.7: 69.4 %, Gemini 3.1 Pro: 68.5 %)

Benchmarks are lab tests. They say little about reality. They also measure the bare model over the API. In ChatGPT, something else arrives: system prompts, safety filters and UI layers change behaviour. ChatGPT is a consumer product, not a business platform. Over the past few months, OpenAI has repeatedly downgraded models in ChatGPT or overwritten them with default prompts, without communicating this transparently. If you want to work with the real model performance, use the API or Codex — not the ChatGPT UI.

Note: Vendor benchmarks from OpenAI itself. What matters is not the table, but how reliably the agent completes a real task without intervention — and how much of that actually reaches the final ChatGPT product.

Sources: OpenAI: Introducing GPT-5.5 | TechCrunch | Decrypt

3. Google: Gemini now creates Word, Excel, PDF and Slides directly in chat

This is new: Gemini can now generate files on request — Word documents, Excel sheets, PDFs, Google Docs, Sheets and Slides. If you need a memo, a table or a slide-deck sketch, you get it straight in the chat as a download. The feature has rolled out globally, applies to all Gemini app users and closes a gap where ChatGPT with Canvas and Claude with artefacts had been ahead until now.

Also in the April drop: native macOS app, personalised image generation, NotebookLM project folders, Lyria 3 Pro music tracks up to 3 minutes free.

The problem is not the model. Google is innovative in parts. But the rollout is chronically slow. Example NotebookLM: project folders are still not available in Switzerland — not even for paying business users. Anyone reading about Gemini features should first check whether they actually reach here. If Google carries on like this, it will fall further behind Anthropic and OpenAI — not technically, but in distribution.

At the same time, on 21 April Google launched Deep Research and Deep Research Max in the Gemini API. Both run on Gemini 3.1 Pro; Max uses more compute for long asynchronous research. Target group: enterprise analysts.

Recommendation: Test file generation in your workflow as soon as it is available in Switzerland for your account. With Gemini features, always check in advance whether they work here too.

Sources: Google: Files in Gemini | Gemini Drop April 2026 | Deep Research Max

4. Mistral releases Medium 3.5 — catches up, small models remain more interesting

Mistral AI published Mistral Medium 3.5 on 29 April 2026: a dense 128B model with a 256K context window, modified MIT licence, available for download on Hugging Face. Score on SWE-Bench Verified: 77.6 %. The model closes the gap to Claude and GPT-5.5 clearly — even if the context window still falls short of 1 million tokens at OpenAI and 200K at Claude.

Mistral's real core strength is not the benchmark, but geography: the datacentre is in France. No US Cloud Act, no Patriot Act, no mandatory data handover to US authorities. For Swiss companies with EU data protection requirements, Mistral is often the only cloud option that avoids third-country transfer questions.

More interesting than the 128B flagship are Mistral's small models. They can run on-premise — on your own GPUs, in your own datacentre, without data leaving the building. 128B needs its own datacentre to run in production; a 7B or 24B runs on four workstation cards. This is where open source becomes usable.

At the same time, Mistral launched Vibe Remote Agents: cloud coding sessions, startable from CLI or Le Chat, with integration into GitHub, Linear, Jira, Sentry, Slack and Teams. As early as 28 April, Workflows came as Public Preview — orchestration with human-in-the-loop and on-prem deployment.

Recommendation: For sensitive data in Switzerland or the EU: check Mistral as a cloud option if US providers are legally problematic. For full control, test the small models on-premise — that is open-source logic in practice.

Sources: Mistral: Vibe Remote Agents + Medium 3.5 | MarkTechPost | The Decoder: Workflows

5. Switzerland & Europe: Apertus 1.5 catches up with Gemma, EU deadlines remain open

Apertus, the open language model from ETH, EPFL and CSCS, gets a 1.5 version of the 8B variant this May. The leap matters: Apertus 1.0 already matched Meta's Llama. Apertus 1.5 now catches up with Google's Gemma models. Switzerland is thus on a par with the US tech giants for the first time in open 8B models. That is something to be proud of.

In practice, this matters for every Swiss company that processes sensitive data and does not want to send it to a US cloud: Apertus runs self-hosted or via Swisscom as a Swiss provider, remains under Swiss law and is fully open and documented. In concrete terms: internal applications with personal data, contracts, patient data or government communications can be built on Apertus without the data leaving the building. The canton of Ticino has been showing this since March for administrative translations.

In Brussels, the EU AI Act deadline remains unclear. In principle, the strict rules for high-risk systems should apply from 2 August 2026. The European Commission wants to postpone this deadline by 16 months; the trilogue ended on 28 April without agreement. For Swiss companies doing business in the EU: keep working towards 2 August 2026. Anyone using AI systems in the EU market should now create an inventory, determine the risk class of each system and close gaps in documentation and risk management.

Swiss law itself will come sector by sector — so no single comprehensive AI regulation, but amendments to individual laws. The federal proposal is expected at the end of 2026.

From the US: Anthropic published election security measures for the 2026 midterms on 24 April. In Anthropic's tests, Opus 4.7 reached 95 %, Sonnet 4.6 96 % on political neutrality.

Recommendation: Keep Apertus 1.5 8B on your radar as soon as it is released. For internal use cases with sensitive data, an open model hosted in Switzerland is often the only clean option.

Sources: Apertus / EPFL | Apertus / Ticino | EU trilogue: DLA Piper | Anthropic: Election Safeguards

Three recommendations for the week

Connector instead of copy-paste: If you have Adobe or Blender in your stack, try the Claude connector in a pilot. Ask: which two or three repetitive steps does it save per project?

Test agentic models with a clear head: GPT-5.5 and Mistral Medium 3.5 promise autonomous multi-step tasks. Rather than trusting benchmarks, run a real workflow — and measure how often human intervention is needed.

Check Apertus 1.5 for sensitive data: As soon as the new 8B model is out, test it with a concrete internal use case. Contracts, HR data, government communications — applications where the data must not leave the country.