The AI Enthusiast.

Claude Opus 4.8 lands — and it catches 4× more bugs than 4.7.

Anthropic shipped Claude Opus 4.8 on May 28, 2026, the follow-up to last month's Opus 4.7. The release is explicitly framed as a reliability and honesty upgrade rather than a raw-capability jump.

The model is the new default for Max, Team Premium, Enterprise pay-as-you-go, and Anthropic API accounts, configured for high effort by default. The headline number from Anthropic's own evaluation: Opus 4.8 is roughly 4× less likely than Opus 4.7 to let flaws in code pass unremarked.

A fast mode variant of Opus 4.8 is also live at $10 input / $50 output per million tokens, putting it in direct price competition with the high tiers of GPT-5.5 and Gemini 3.5 Pro. For users on Claude Code, the swap is silent — you're already running 4.8 on every new session unless you've pinned a model.

Anthropic files for IPO at a $965B valuation.

On Monday, June 1, 2026, Anthropic confidentially filed for an IPO, capping a funding round that values the company at $965 billion.

The round itself raised $65 billion in fresh capital, much of which is earmarked for the compute deals quietly piling up — SpaceX/Colossus, the announced data-center expansions, and now reportedly a slice of SoftBank's €75B France build.

Reading between the lines: Anthropic is going public roughly two years earlier than expected. CBS News framed it as "a public-market test of the AI boom," and that's not hyperbole. There has been no comparable pure-play AI IPO at this scale. Whether the S-1 reveals enterprise revenue that justifies the price, or shows the lopsided spend on training and compute that everyone suspects, will set the tone for the entire sector through Q3.

Claude Code goes cloud-native: ultraplan, ultrareview, dynamic workflows.

Three Claude Code features shipped in the last three weeks fundamentally change what "running Claude Code" means.

• Ultraplan (early preview) — draft a plan in the cloud from your CLI, review and comment on it in a web editor, then either run it remotely or pull it back local. The local-only mental model is officially gone.
• claude ultrareview — a cloud fleet of bug-hunting agents you can drop into CI pipelines or scripts. Multi-agent code review without you wiring up the orchestration.
• Dynamic workflows — orchestrate numerous sub-agents from a single script, with the runtime managing their context and hand-offs.

A new security-guidance plugin also landed that proactively flags vulnerabilities in real-time as Claude modifies code. And on Windows, Git Bash is no longer required — Claude Code defaults to PowerShell. Auto mode reached the Pro plan and now supports Sonnet 4.6 as well.

Google ships Antigravity Agent — Gemini gets a sandbox.

At I/O the throughline was "agents." In the three weeks since, Google operationalized it.

Gemini 3.5 Flash went GA as Google's "most intelligent model for sustained frontier performance on agentic and coding tasks." More importantly, two agent primitives landed in the Gemini API:

• Managed Agents (public preview) — build and deploy autonomous, stateful agents that run inside Google-hosted Linux sandbox environments. Direct competitor to Claude Managed Agents.
• Antigravity Agent (public preview) — a managed agent that can autonomously plan, reason, write and execute code, manage files, and browse the web inside its sandbox container. Google's pitch is parity-or-better with OpenAI's Operator and Anthropic's Claude Code background agents.

File Search also gained multimodal support — embed and search images natively using Gemini embeddings. That's the kind of small platform feature that quietly removes 200 lines of RAG glue from your codebase.

Gemini Spark, Daily Brief, Omni video — Google's consumer push.

Hidden behind the developer announcements, the Gemini app got the biggest consumer overhaul since launch.

• Daily Brief — a personalized morning summary built from your Gmail, Calendar, Drive, and recent activity. The pitch is "an AI that knows what your day looks like before you open your laptop."
• Gemini Spark — a personal AI agent that lives inside the app, persists context across conversations, and can take action on your behalf (book, draft, schedule, summarize).
• Gemini Omni — a new AI video model exposed inside the Gemini app, joining Veo as the consumer-facing video generation surface.
• Redesigned interface — Google has finally shipped a Gemini UI that doesn't feel like a Bard skin.

For non-developer users, this is the moment Gemini went from "ChatGPT alternative" to a credibly different product — built around proactive briefings and personal-agent workflows that ChatGPT still doesn't quite have.

OpenAI ships Dreaming V3 memory — and a ChatGPT ad platform.

Two big OpenAI moves in the last week.

• Dreaming V3 — a "significant memory upgrade" began rolling out to ChatGPT Plus and Pro users in the US on June 4, 2026. Early reports describe it as multi-thread, multi-month context that ChatGPT actually uses unprompted — a sharp upgrade from the existing Memory feature, which still has to be reminded what's relevant.
• ChatGPT self-serve advertising platform — OpenAI is opening up paid placements inside ChatGPT for brands. This is the first time OpenAI has shipped an ad product, and the model still has to credibly disclose paid content without nuking trust.

OpenAI also launched real-time audio and translation models for agents, making voice agents finally viable for support, sales, and education. Separately, it gave the EU access to GPT-5.5-Cyber, a variant for vetted cybersecurity teams, EU businesses, governments, and EU institutions.

MCP hits 97M downloads — and some teams are already leaving.

The Model Context Protocol crossed 10,000+ servers and 97 million monthly SDK downloads, with every major lab now supporting it.

Anthropic in particular leaned in this week: 20+ new legal MCP connectors and 12 practice-area plugins shipped for law firms, and Claude Managed Agents can now connect to your private MCP servers inside a sandbox you control.

The counter-trend, though, is real and growing: more teams are moving away from MCPs entirely for production pipelines, going back to direct API calls and CLIs. The reason is token economics — MCP schema loading and protocol overhead eat context budget, and in production token efficiency wins over convenience. MCP's sweet spot has clarified: auth, multi-tenancy, enterprise governance, and non-technical teams. Production data pipelines, often not.

Meanwhile, the A2A (Agent-to-Agent) protocol — Google Cloud's horizontal agent-comms standard, now governed by the Linux Foundation — has hit 50+ technology partners, positioning itself as the layer above MCP rather than a replacement.

Project Glasswing expands — Claude Security goes GA-ish for 150 new orgs.

Anthropic expanded Project Glasswing, its enterprise rollout program, granting Claude Mythos Preview access to roughly 150 new organizations.

Bundled with the expansion: Claude Security, which performs codebase scans and suggests patches automatically. The combination matters because it puts Anthropic into direct competition with GitHub Advanced Security and Snyk — but with the model that wrote the code also reasoning about the vulnerability.

Patch suggestions aren't theoretical fixes; they're concrete diffs the same Opus 4.8 generation could land in CI. Combined with the new security-guidance plugin in Claude Code (Story 3), Anthropic is methodically building the "trust layer" enterprises have been asking for since the agent era began.

EU AI Act gets a quiet delay — and two new prohibitions.

On May 7, 2026, EU negotiators reached a provisional agreement on the Digital Omnibus on AI — the first set of amendments to the EU AI Act since it was adopted in June 2024.

Delays

• High-risk AI obligations under Annex III are postponed from August 2, 2026 to December 2, 2027 — a 16-month deferral.
• Member-state regulatory sandbox obligations slip from August 2026 to August 2027.

New prohibitions (effective December 2, 2026)

• AI systems generating or manipulating non-consensual intimate material.
• AI systems generating child sexual abuse material (CSAM).

The delays reflect the very real challenge of operationalizing third-party assessments and conformity testing for high-risk AI — the requirements simply weren't deployable on the original schedule. The new prohibitions are a long-overdue and uncontroversial addition. Final adoption is expected in late June 2026 with publication in July.

The compute buildout goes vertical: €75B France, $1T SK Hynix.

Two numbers from this window that lock in the "AI infra is the real story" thesis.

• SoftBank committed up to €75 billion to build 5 gigawatts of AI data center capacity in France, with a first phase of €45B delivering 3.1 GW in Hauts-de-France by 2031. This is roughly the size of the entire current French data-center market built again — and oriented entirely toward AI workloads.
• SK Hynix crossed a $1 trillion market cap for the first time, joining Samsung and Micron as the AI-memory boom remakes the semiconductor industry. HBM (high-bandwidth memory) demand is now the single largest driver of semi capex.

Layer in NVIDIA Cosmos 3 + Intel Xeon 6+ shipping faster, cheaper AI processing this month, and a White House executive order on "Promoting Advanced AI Innovation and Security" issued in June, and the picture is clear: AI's bottleneck is no longer model capability or talent — it's electrons, copper, and HBM stacks.