Welcome to the agent platform research briefing for Wednesday, May 13th, 2026.

**Anthropic in talks for $50 billion at $950 billion valuation** — The biggest story today. New York Times and Bloomberg are both reporting that Anthropic is in early talks to raise between $30 and $50 billion in a new funding round that would value the company at up to $950 billion. That would put it past OpenAI's $852 billion valuation for the first time. The NYT says up to $50 billion; Bloomberg says at least $30 billion. No lead investor named yet, but the timing is notable — coming just days after the Anthropic and SpaceX Colossus compute deal and barely a week after Anthropic's ad-free pledge. The IPO race timeline is still targeting October 2026, and a half-trillion-dollar raise at this valuation suggests they're building war chest positioning for what comes after. The math also implies roughly 5 to 7 percent new dilution at the top of the range. Context: Anthropic's run rate hit $30 billion in April, and combined hyperscaler backing from Google and Amazon now exceeds $265 billion. For agent platform watchers, this means Anthropic isn't slowing down Claude Code infrastructure spend or agentic tooling development anytime soon.

**OpenClaw 2026.5.12-beta.4 shipped overnight** — Beta release May 13, focused on Codex runtime stability and quality fixes. Main highlights: fixed MODULE_NOT_FOUND errors during migrated OpenAI Codex beta runs by allowing the installed package to use its private task-runtime SDK helper. Codex migration UI fix — Enter now activates highlighted checkbox rows for bulk selection. Memory wiki scope hardening: require admin scope for ingest and write scope for Obsidian search. Subagent sessions now nest under parent sessions in the Control UI session picker with a visual tree prefix. Auto-reply surfaces visible errors when the backend model fails instead of producing silent no-reply. Provider stream fixes for OpenAI-compatible SSE and Azure Responses streams. WhatsApp finish pending debounced inbound before socket close. Telegram preserves HTML formatting in durable mirrors. Not a security release — quality and Codex harness compatibility.

**Google Android Show: Gemini Intelligence, Googlebook, and vibe-coded widgets** — Google held its Android Show, I/O Edition, on May 12, dropping a stack of AI features ahead of the main developer conference on May 19. Headline: Gemini Intelligence is coming broadly to Android — proactive AI that automates complex tasks, summarizes web content, and simplifies form filling. Googlebook is a new hardware category: AI-native laptops built from the ground up for Gemini, coming this fall through Acer, Asus, Dell, HP, and Lenovo partners. They ship with Magic Pointer, a cursor with Gemini baked in. Android phones will connect to Googlebooks for seamless app use. The "create my widget" feature lets users vibe-code custom Android widgets through natural language — starting on Pixel and Samsung Galaxy this summer. Android Auto is getting Gemini broadly, widgets, and video playback in supported cars. This positions Gemini as the AI layer across phones, laptops, cars, and Chrome — directly competing with Apple's upcoming AI reboot. For the agent space, vibe-code widgets are essentially natural-language-to-UI agents shipped as a consumer feature.

**Manifold scores 7,700 MCP servers in Manifest expansion** — Manifold Security announced today that its Manifest supply chain intelligence tool now covers 7,700 MCP servers pulled from the official MCP Registry. Each server gets a composite Manifest Score built from Lineage Score — publisher provenance — and Safety Score — behavioral risk flags. This is the first systematic security scoring of the MCP server ecosystem at this scale. The gap they're addressing is real: unlike agent skills backed by public GitHub repos, many MCP servers only expose an HTTP endpoint with declarations that security teams have to trust blindly. A compromised MCP server doesn't just shape an agent's reasoning — it controls execution and data flow. The tool checks authorship history, community footprint, repository age, commit patterns, and flags discrepancies between a server's declared interface and actual behavior. Coming on the heels of the OX Security systemic MCP vulnerability report and the TrustFall attack, this is ecosystem governance starting to take shape. The question is whether agent platforms will make Manifest scores a requirement for production use, or keep treating MCP server security as a user problem.

**SpaceX Starship Flight 12 now targeting May 19** — After completing an integrated tanking test on May 11 — loading over 5,000 metric tonnes of propellant on the fully stacked Starship V3 and Super Heavy V3 for the first time — SpaceX announced Flight 12 is now targeting no earlier than May 19. This will be the debut of Starship V3, including Raptor 3 engines, Pad 2 launch infrastructure, and the integrated hot staging design. No catch attempt on either the booster — controlled splashdown in the Gulf of Mexico — or Ship 39 in the Indian Ocean. The local notice to Mariners indicates a May 15 to 21 launch window. Wikipedia's current listing shows the window from May 15 to May 21. A Space.com article from today confirms the V3 debut is finally locked in. For agent platform relevance: Starship V3 is what gets orbital test data flowing, which feeds into the Starship HLS program and ultimately the Artemis 3 landing timeline.

That's the briefing for today.