AI Weekly Briefing: 23-29 May 2026

Anthropic dominated the week, with investors pricing Claude like critical infrastructure while the rest of the market pushed agents further into finance, operating systems, enterprise software delivery and media workflows.

Anthropic reaches a $965bn valuation

Anthropic raised $65bn in private funding, pushing its valuation to $965bn, according to reports from Al Jazeera, AP and The Guardian published on 28-29 May. The round puts the five-year-old Claude maker ahead of OpenAI by reported valuation and places it within touching distance of the trillion-dollar club before any expected IPO.

The funding story is inseparable from Claude’s enterprise demand. Coverage this week pointed to large cloud and compute relationships across Amazon, Google, Microsoft and SpaceX, while the company’s own product cadence is increasingly centred on coding agents. The number is extreme, but investors appear to be underwriting Anthropic as a full-stack AI platform: model lab, enterprise vendor, developer-tool supplier and future public-market candidate.

Claude Opus 4.8 arrives with Dynamic Workflows

Anthropic also released Claude Opus 4.8 on 28 May, with TechCrunch reporting a new Dynamic Workflows capability for coordinating swarms of subagents. The company framed the model as a “modest but tangible improvement” over its predecessor, but the product emphasis is clear: Claude is being pushed deeper into codebase-scale work rather than simple chat or file-by-file assistance.

The most concrete claim is that Claude Code with Opus 4.8 can handle migrations across hundreds of thousands of lines of code, using the existing test suite as its bar. That is the right benchmark for enterprise buyers. Teams do not need another demo where an agent builds a toy app. They need migration, remediation and test-driven change across real repositories.

Robinhood gives AI agents a route into trading

Robinhood launched support for AI agents that can analyse portfolios, suggest trading strategies and place orders using funds held in a dedicated wallet, according to TechCrunch, Fortune, Forbes and Proactive Investors coverage from 27-28 May. Users can pause autonomous trading and preview agent-made trades, while access is initially aimed at technically confident early adopters.

That makes the launch one of the clearer tests of agentic AI in a regulated, consumer-facing financial product. The dedicated wallet limit matters because it creates a boundary around agent action, but the underlying move is still significant. An assistant that can talk about stocks is one thing. An agent with delegated execution authority, even inside constraints, is a different risk and product category.

Google and the industry confront agent security in real time

TechCrunch’s 24 May security piece captured the week’s awkward transition: large companies want agents inside real workflows, but the security model is still catching up. The warning was direct. Agents moving through an enterprise will find data assets and may expose data that humans did not realise was reachable.

Google’s position in that story is useful because it shows the problem is not confined to small vendors shipping quickly. Even the platform companies are still working through how permissions, observability, policy enforcement and machine-speed defence should work when agents can discover systems, call tools and chain actions. Security teams will need controls designed for autonomous software behaviour, not only user sessions and API keys.

OpenAI turns Codex into an enterprise case-study engine

OpenAI published a run of Codex customer stories this week, including Endava on 28 May and Virgin Atlantic on 22-28 May. Endava says it is using Codex to reduce requirements analysis from weeks to hours, while Virgin Atlantic credits Codex with helping ship a revamped mobile app against a fixed holiday travel deadline, reaching near-total unit test coverage and zero P1 defects.

Those are vendor case studies, so the claims should be read with the usual caution. Still, the pattern is worth noting. OpenAI is no longer selling coding agents only as developer productivity tools. It is positioning Codex as a delivery-system change for large organisations, with analysis, design, implementation and testing pulled closer together under agent-assisted workflows.

Microsoft keeps shaping Windows around agents

Coverage of Build 2026 this week described Microsoft positioning Windows as an agent platform for AI developers. That continues a wider shift from Copilot as a feature layered onto Windows towards Windows as an environment where agents can run, observe context and connect to developer workflows.

For builders, the operating-system angle matters because agents need durable context, permissions and user trust. Browser tabs and chat windows can carry early adoption, but deep agent work eventually runs into platform boundaries. Microsoft appears determined to make Windows one of the places where those boundaries are redrawn rather than merely patched around.

Also worth noting

OpenAI published its Frontier Governance Framework on 28 May, tying its safety, security and risk practices to emerging EU and California regulatory requirements. The timing matters because frontier labs are trying to show that they can formalise oversight before legislators define every operational detail for them.

Runway introduced Runway MCP on 27 May, giving agents and coding tools a Model Context Protocol route into its image and video generation tools. BCD also announced MCP use across its Tripsource platform, while MCP community activity this week pointed to the protocol’s widening role as the connective tissue for agent products.

WIRED published a 26 May feature on the rise and disorder of AI agents, including OpenClaw as part of the broader personal-agent wave. It is a useful counterweight to vendor launches: the same week that agents moved into finance and enterprise delivery, the cultural story remained one of speed, experimentation and unresolved norms.