This was a workflow week.

Anthropic shipped Claude Opus 4.7, framing it as a stronger model for coding, agents, vision, and longer multi-step work. OpenAI launched GPT-Rosalind, a purpose-built model for biology, drug discovery, genomics, and translational medicine. Google kept expanding Gemini’s surface area with personalized image creation, interactive simulations, and Skills in Chrome, which turns frequently used prompts into reusable one-click workflows.

The bigger pattern: we are moving past the phase where every week is just “new model, better benchmark.” The real fight now is over who owns the daily workflow — design, coding, research, note-taking, analysis, and task automation. That is where users stick, and where revenue compounds. This is an inference drawn from the week’s product launches and company positioning.

Anthropic released Claude Opus 4.7 on April 16 and described it as generally available, with improvements in software engineering, agents, vision, and multi-step task performance. Anthropic also says it sees images in higher resolution and is more consistent on complex professional work.

Why this one matters: Claude keeps tightening its grip on the “serious work” crowd. This was not pitched as a flashy consumer feature drop. It was pitched as a better engine for people doing technical, long-horizon, high-context tasks. That is a very valuable lane in this market. This interpretation is based on Anthropic’s own product framing and rollout notes.

OpenAI introduced GPT-Rosalind on April 16 as a frontier reasoning model built for life sciences research, with emphasis on chemistry, protein engineering, genomics, drug discovery, and translational medicine.

This is one of the clearest signs yet that vertical AI is becoming a real product category, not just a pitch deck idea. Instead of trying to make one general model feel good enough for every industry, OpenAI is showing what happens when a top lab packages capability around a specific high-value domain. Expect more of this across law, finance, logistics, and security. The last sentence is an inference from the launch and broader enterprise packaging push.

Google introduced new ways for the Gemini app to generate personalized images using users’ interests, preferences, and Google Photos context, and it also rolled out Gemini’s ability to create interactive visualizations and simulations directly in chat.

That combination matters because it shows Google pushing on two different fronts at once: personal context on one side, useful output formats on the other. One makes Gemini feel more tailored. The other makes it more useful for explaining ideas, learning concepts, and presenting information.

Google launched Skills in Chrome on April 14, letting users discover, save, remix, and instantly repeat AI workflows built from prompts.

This is the kind of feature that sounds small until you use it. A saved prompt is nice. A reusable workflow is better. Products get a lot stickier when users stop asking “what should I prompt?” and start clicking a repeatable action that already works. That last point is an inference from the product’s stated purpose and standard workflow-retention dynamics.

OpenAI also launched Codex for (almost) everything, expanding the Codex app with computer use, app connections, image generation, memory, browser support, remote devbox access over SSH, PR review, and multi-terminal workflows.

Google added notebooks in Gemini, linking the Gemini app more tightly with NotebookLM so users can keep project materials, chats, and files organized in one place.

On the open side, Google’s Gemma line kept gaining momentum: Google’s developer documentation shows Gemma 4 releases on March 31, and the Gemma landing page positions the family for advanced reasoning and agentic workflows.

The throughline across all of that is pretty clear: the winners are trying to become systems of work, not just assistants you visit occasionally. That is an inference from the product changes across Anthropic, OpenAI, and Google this week.

Perplexity
Perplexity has been expanding well beyond AI search. Recent product updates include Comet on iOS, Computer updates for both individuals and enterprise users, multimodal Deep Research, and Health Computer. The broader story is that Perplexity is moving toward an AI-native browser and workflow layer, not just an answer engine. (perplexity.ai)

Google AI
Google had a broader AI week than just Gemini app features. The biggest miss was Gemini Robotics-ER 1.6 from DeepMind, which Google says improves embodied reasoning, multi-view understanding, and instrument reading. On top of that, Google rolled out interactive simulations and models in Gemini, notebooks in Gemini, personalized image creation, and new Gemini API inference tiers. The overall theme was Google pushing simultaneously on consumer AI, developer infrastructure, and robotics. (deepmind.google)

OpenClaw
OpenClaw is one of the more notable open-agent names to watch right now. Its GitHub releases show active April development, including 2026.4.15-beta.1, and the pace of shipping suggests real momentum in the open agent ecosystem. It matters less as a mainstream consumer brand and more as a sign that open-source agent frameworks are maturing quickly and influencing how larger companies think about autonomous workflows.

A very usable setup right now looks like this:

Why this stack works: each tool is leaning into a different part of the workflow instead of trying to be everything at once. Claude is strongest when the work is messy and thoughtful. Codex is strongest when you want execution. Gemini’s notebook layer is useful when the work starts turning into a project instead of a chat. Chrome Skills helps remove prompt friction. This is an editorial synthesis based on the product descriptions.

If you want something less dependent on one closed frontier provider, keep an eye on Gemma 4 plus the surrounding open tooling ecosystem. Google’s docs position Gemma 4 for advanced reasoning and agentic workflows, and Hugging Face highlighted both Gemma 4 deployment support and new posts on multimodal embedding/reranker training this week.

That does not mean open has “won.” It does mean the open lane is getting more practical for teams that care about portability, customization, or cost control. That conclusion is an inference from the release support and documentation footprint.

Broadcom remains one of the clearest public-market AI beneficiaries this week. Reuters reported that Broadcom signed a long-term agreement with Google through 2031 to develop and supply custom AI chips and components for next-generation AI racks, and that Anthropic will get access to roughly 3.5 gigawatts of AI compute capacity drawing on Google processors starting in 2027.

On top of that, Reuters also reported that Meta extended its custom-chip partnership with Broadcom through 2029, with an initial phase representing more than one gigawatt of compute capacity.

The story here is simple: Broadcom keeps showing up wherever hyperscalers want custom AI silicon and networking at scale. That is why it belongs near the top of any “AI infrastructure beneficiaries” watchlist. The second sentence is an inference based on the Google, Anthropic, and Meta deals.

Google’s AI story this week had two layers: consumer product velocity through Gemini, and deeper infrastructure leverage through custom silicon. The Gemini-side launches included personalized image generation and interactive simulations, while Reuters reported the long-term Broadcom chip agreement through 2031.

That mix matters because Google is one of the few companies that can compete across models, distribution, cloud, and chips at the same time.

Meta’s catalyst this week was more infrastructure than app-demo. Reuters reported that Meta extended its Broadcom custom-chip partnership through 2029 to support AI ambitions across its platforms.

Meta is still one of the most important AI distribution stories in the market, but custom silicon is becoming a bigger part of the thesis because it supports inference economics and scale across its product ecosystem. The second clause is an inference from the reported partnership expansion.