The AI Watch Returns

Five weeks since the last edition. In that window, the AI ecosystem lived through two of its busiest weeks of the year. Anthropic shipped Claude Opus 4.7. OpenAI answered with GPT-5.5. Google announced the largest AI deal in history — $40 billion into Anthropic. And the Vertex AI platform died, replaced by the Gemini Enterprise Agent Platform.

I had promised you a weekly watch, and I missed several appointments. My bad. Rather than drowning you in five weeks of news, I'm focusing on what matters: the past two weeks (April 18 → May 2). No hype, no buzzwords, just what concretely changes for your workflows and tech decisions.

The Major News of the Past Two Weeks

Google invests up to $40 billion in Anthropic — April 24: It's the largest AI deal ever closed. Google injects $10 billion immediately at a $350 billion valuation, with additional tranches gated by milestones, up to $40 billion total. More importantly, Google Cloud is providing 5 gigawatts of compute over 5 years — the energy equivalent of a major metropolitan area dedicated to training models. The strategic signal is clear: Google can no longer compete head-on with OpenAI on consumer, so it's funding the only credible enterprise rival. For you: in 12 months, only 3 or 4 real foundational model players will remain. Prepare for an ultra-concentrated market and architect your workflows to avoid single-vendor lock-in.

Claude Opus 4.7 — April 16: Anthropic doesn't do hype anymore — just good models. Opus 4.7 jumps to 87.6% on SWE-bench Verified (vs 80.8% for 4.6) and introduces "task budgets" — you give Claude a token budget for an agentic loop, and it prioritizes its work based on the remaining countdown. Also the first Claude model to support high-resolution images (up to 2576px). I switched on day one. The difference shows mostly on long tasks: the model "holds" context better over 30 minutes of autonomous agent work, where 4.6 used to drift. Pricing unchanged at $5/M input, $25/M output.

GPT-5.5 — April 23: OpenAI counters Opus 4.7 within a week. SWE-bench at 88.7% (slightly above Claude), -60% hallucinations vs GPT-5.4. The model is unified: a real-time router picks between the standard version and GPT-5.5 Thinking based on query complexity. Concretely, you stop having to choose between speed and reasoning — the API does it for you. It's also the first OpenAI model genuinely reliable for long agent workflows, closing a real gap with Anthropic. Worth testing in parallel with Claude to see which performs better on your specific use cases.

Google Cloud Next '26 — Gemini Enterprise Agent Platform — April 22: Vertex AI is dead, long live the Gemini Enterprise Agent Platform. Google unified its entire AI agent stack into a single platform: Agent Studio (low-code visual), Agent Development Kit (code-first), Model Garden (200+ models including Gemini 3.1 Pro), and a graph-based framework to orchestrate sub-agents. For companies already on Google Cloud, this is a game changer — you can now build enterprise agents without gluing five different tools together. For everyone else, the question becomes: stay on AWS Bedrock or migrate to Google?

DeepSeek V4 — April 24: The Chinese release V4 in open-weight preview, MIT license. Two variants: V4-Pro with 1.6 trillion total parameters (49 billion active), and V4-Flash lighter (284 billion / 13 billion active). 73% fewer FLOPs than V3 for equivalent performance. If you have sovereignty or cost constraints, this is seriously worth looking at. MIT license, runs on accessible hardware (V4-Flash on 4 A100s), and avoids US-provider dependency. For European SaaS and public sector, this might be the most important release of the fortnight.

New Models to Know

gpt-image-2 — April 21: OpenAI's first image generation model with chain-of-thought reasoning built into the creation process. Concretely, the model "thinks" before generating, producing coherent images even on complex prompts: legible typography, multi-object composition, scenes with causality. For marketers producing visuals at volume, it's the best generator available today — more reliable than Midjourney on business constraints (logos, brand consistency, text integration).

Anthropic Managed Agents — beta launched April 8: Anthropic now offers to host your AI agents for you, at $0.08 per session-hour. You pay for active sessions, not tokens. For agents running in the background for hours, it's radically cheaper than a typical API bill. And Anthropic handles infra, state persistence, and sandboxing security. If you're building a product based on long-running agents (which I'm currently testing on Optimzd), put it in your POC immediately.

OpenAI Agents SDK update — April 15: Major SDK update with a model-native harness, reinforced safety controls, and better debugging tools. For Python or TypeScript developers in the OpenAI ecosystem, this is the stack to use. The model-native harness is interesting: OpenAI stops treating tools as an add-on and integrates them directly into the model, reducing latency and improving call reliability.

AI Tools to Activate This Week

Adobe CX Enterprise — April 20: Adobe launches an end-to-end agentic system to orchestrate the customer lifecycle. The clever bit: Adobe CX Enterprise Coworker executes tasks defined by business goals, not static rules. You tell it "increase LTV for segment X clients," and it coordinates multiple sub-agents (segmentation, content, distribution, measurement). For companies already on the Adobe stack (Marketo, Experience Cloud), it's a natural upgrade. For others, worth seeing what it does before committing to HubSpot Breeze or Salesforce Agentforce.

OpenAI Workspace Agents: ChatGPT for Business and Enterprise can now create and share agents that autonomously execute tasks across Slack, Gmail, and other integrated tools. Concretely, your team can build an agent that "summarizes important Slack threads each morning" or "classifies incoming emails" without writing a line of code. It's the direct competitor to Microsoft Copilot Studio, with better UX and the ChatGPT ecosystem behind it. If you're already on ChatGPT for Business, activate this before considering other tools.

Gemini 3.1 Flash Image: Mentioned in passing at Cloud Next, it's the direct competitor to gpt-image-2. Faster, cheaper, slightly behind on coherence for complex prompts. For high-volume image generation (e-commerce, social media, product catalog), it's probably the best price/performance ratio currently.

The Numbers That Matter

$40 billion — Google's total investment in Anthropic. Compare that to OpenAI's $110 billion total raised since 2019. The foundational model market reaches a capitalization level only 5 or 6 sectors in the world have ever seen.

5 gigawatts — the compute capacity Google Cloud will dedicate to Anthropic over 5 years. The electrical equivalent of several nuclear plants. Compute has become the new barrier to entry in the AI sector.

$350 billion — Anthropic's valuation at this round. Anthropic is now worth more than Disney, Volkswagen, or Coca-Cola. For a 4-year-old company with fewer than 1,500 employees, this is unprecedented.

88.7% vs 87.6% — SWE-bench Verified scores for GPT-5.5 and Claude Opus 4.7. For the first time in 18 months, OpenAI moves back ahead of Anthropic on the coding benchmark. But the gap is marginal — model choice now depends mostly on specific needs (latency, context, cost, integration).

20,000 jobs cut — Meta and Microsoft jointly announced over 20,000 layoffs on April 24. Over 92,000 tech jobs eliminated in 2026 so far. The official narrative: "reorienting toward AI." The real narrative: AI is effectively replacing middle-office and engineering functions faster than expected.

40% — the share of enterprise applications that will embed an AI agent by end of 2026, according to Gartner. But only 10% of organizations have actually scaled their agents — most stay stuck in pilot mode due to governance issues, not tech ones.

My Take — What I'm Actually Doing With This

After testing most of these releases, here are my pragmatic recommendations:

For code and long agents → I switched to Claude Opus 4.7 on day one and haven't looked back. Task budgets change the equation for agentic workflows over 10 minutes. If you were on Sonnet 4.6 to save money, the Opus 4.7 premium is justified on complex tasks — the increased reliability more than compensates for the API bill.

For enterprise agent infrastructure → If you're starting today, choosing between Gemini Enterprise Agent Platform, OpenAI Agents SDK and Anthropic Managed Agents is the most important tech decision of the year. My bias: Anthropic Managed Agents for the per-session pricing, unless you're already deep in Google Cloud. For simple workflows, OpenAI Workspace Agents gets the job done without code.

For image generation → gpt-image-2 for critical visuals (campaigns, hero images, brand content), Gemini 3.1 Flash Image for volume (product catalog, social media). Midjourney remains best for "artistic" style, but for business content, these two models are superior on reliability.

For model strategy → The Google→Anthropic deal confirms what I've been saying for 6 months: we're heading toward an OpenAI/Anthropic duopoly, with Google and Meta as outsiders. Don't build your stack around a single model. Architect your workflows to easily switch between Claude, GPT, and an open-source (DeepSeek V4, Llama 4) depending on context. A homemade router or a tool like OpenRouter is 2 days of dev that saves you from rebuilding your entire integration in 6 months.

For sovereignty → DeepSeek V4 under MIT license is serious. If you have strict GDPR constraints or sensitive data, V4-Flash runs on 4 A100s and rivals GPT-5.4 on most benchmarks. For European SaaS and public sector companies, it's worth an immediate POC — you save on API costs and keep your data in-house.

See You Next Week

Promised, I'm sticking to it this time. The cadence returns to weekly: no hype, no buzzwords, just what you can activate Monday morning.

If a topic interests you in particular — the Anthropic deal, Managed Agents, or arbitraging between the three enterprise platforms — let's talk. See you next week.