Anthropic Closes $65B Series H at a $965B Valuation, Eclipsing OpenAI
Anthropic disclosed a $65 billion Series H round at a $965 billion post-money valuation, co-led by Altimeter Capital, Dragoneer, Greenoaks, and Sequoia, with Samsung, SK Hynix, and Micron joining as “strategic infrastructure partners.” The round vaults the five-year-old company past OpenAI to become the world’s most valuable private startup — and lands as Anthropic preps for a likely Wall Street debut.
The fundraise is paired with eye-popping operating numbers: run-rate revenue reportedly crossed $47 billion this month, and Anthropic is forecasting a 130% surge that would push it to its first operating profit. The chip-maker triad is the real strategic tell. Samsung — the only one of the three with foundry capacity — is already shipping samples of its 12-layer HBM4E memory built for AI servers, and the partnership opens a path to custom logic silicon down the road.
Claude Opus 4.8 Lands: 69.2% SWE-Bench Pro, Dynamic Workflows, 2.5x Fast Mode
Anthropic released Claude Opus 4.8 less than two months after Opus 4.7. The headline numbers: agentic coding (SWE-Bench Pro) jumps from 64.3% to 69.2%, multidisciplinary reasoning with tools moves from 54.7% to 57.9%, and agentic computer use climbs from 82.8% to 83.4%. Opus 4.8 also flags about 4x fewer code flaws than Opus 4.7 and beats GPT-5.5 (58.6%) and Gemini 3.1 Pro (54.2%) on the same coding bench.
Pricing is unchanged at $5 per million input and $25 per million output tokens, with a 1M-token context window on the Claude API, Bedrock, and Vertex AI. The bigger product story is Dynamic Workflows in research preview: ask Claude to build a workflow and it orchestrates tens to hundreds of background agents to deliver the result. Fast Mode now runs 2.5x faster at the same price — and roughly 3x cheaper than the equivalent on prior models.
Anthropic Re-Commits Claude to Ad-Free
Anthropic took the unusual step of putting its no-ads stance in writing this week: Claude products will not carry advertising, and the company will not let advertisers pay for placement inside conversations. The framing: ad incentives are incompatible with a genuinely helpful assistant, because the moment ranking is for sale, recommendations stop being for the user.
It’s a positioning move with real teeth. Anthropic just raised $65B at a near-trillion valuation on enterprise economics, not consumer ad inventory, so this is an easy commitment to keep — but it also forces a contrast with peers leaning into commerce, search ads, and sponsored content. Worth noting: the policy applies to Anthropic’s own products; developers building on the API can still serve their own ads.
Anthropic Acquires Stainless to Own the SDK and MCP Server Layer
Anthropic completed its acquisition of Stainless, the SDK-generator and MCP server tooling company used by hundreds of API providers to ship libraries, CLIs, and connectors. Stainless customers include some of the largest names in API land — bringing the team in-house gives Anthropic ownership of the layer that turns an API into a usable agent surface.
The strategic logic is obvious if you’ve been watching the MCP build-out. Anthropic donated the protocol to the Agentic AI Foundation in December, then quietly invested in the connector tooling that lets every SaaS vendor become a Claude target. Owning Stainless tightens that supply chain and pulls a friction-removing step closer to the agent loop, with predictable downstream effects on every developer relations strategy in the model space.
GitHub Copilot Makes Claude Opus 4.8 Generally Available
GitHub flipped Claude Opus 4.8 to GA across Copilot the same day Anthropic launched the model. The integration covers Copilot Chat, agent mode, and Copilot Coding Agent — meaning the same model that posts a 69.2% SWE-Bench Pro is now selectable inside the dev tool with the largest active install base.
Coming a few days after the Claude Code team added /effort xhigh for the hardest tasks and a security plugin for inline vulnerability flags, this is a useful pincer for Anthropic: keep Claude Code the lead surface for power users while making sure Opus 4.8 is one click away in Copilot for everyone else. The pricing arbitrage matters too — Opus 4.8 Fast Mode is roughly 3x cheaper than Fast Mode on earlier models, which is the kind of detail that quietly shifts default model selection inside engineering orgs.
System Messages Mid-Conversation: Cache-Safe Instruction Updates
The Messages API now accepts system entries inside the messages array, letting developers update Claude’s instructions mid-task without breaking the prompt cache or routing the change through a synthetic user turn. Small change, real impact: long-running agent loops can now adjust persona, tool permissions, or safety constraints between turns without paying the re-prefill tax.
Pair this with the new self-hosted sandboxes (public beta) and MCP tunnels (research preview), and the agentic surface area for enterprise builders gets a lot more flexible this week. Skills (beta), the Files API (beta), and the MCP connector also remain available alongside batch processing and prompt caching for cost control.
Samsung’s HBM4E Sample Shipments Tie Directly Into the Anthropic Round
Samsung disclosed it has begun shipping samples of its 12-layer HBM4E chip — high-bandwidth memory built specifically for AI servers — right as it joined Anthropic’s Series H. SK Hynix and Micron, the two other memory giants in the round, were named alongside Samsung as “critical to the world’s supply of memory, storage, and logic chips.”
The read-through for the rest of the AI stack is straightforward. Memory is the binding constraint on inference economics for frontier-class models, and Anthropic just bought itself preferred-customer access at three of the four shops that can supply it. The longer game — given Samsung is the only one with foundry capacity — is custom logic silicon, which would put Anthropic in the same neighborhood as OpenAI’s reported Broadcom partnership.
MCP Apps Framework Lands With OpenAI, Slack, Asana, Canva, Figma, Box
The MCP Apps extension — co-developed by Anthropic and OpenAI, with launch partners including Amplitude, Asana, Box, Canva, Figma, Slack, and Salesforce — brings interactive, app-like surfaces into chat clients. MCP servers can now render UI components inside Claude, ChatGPT, Goose, and VS Code, turning “tools” into something closer to embedded mini-apps.
That this shipped jointly with OpenAI is the news. After Anthropic donated MCP to the Agentic AI Foundation last December, the protocol’s neutrality has been an open question. The MCP Apps launch — cross-vendor, cross-client — is the strongest signal yet that MCP is settling into the role of universal agent connector, regardless of which model is doing the reasoning.
The Capital and the Cadence Are Now the Story
Two numbers from this week tell you almost everything. $65B at a $965B valuation. Opus 4.7 to Opus 4.8 in under two months. Anthropic is now raising at a scale OpenAI is unaccustomed to being underbid on, and shipping at a tempo that makes the “flagship model release” feel like a quarterly event rather than an annual one. Wrap those two signals in the rest of May — KPMG, PwC, SAP, BMS, Salesforce, AWS, 28 security integrations, MCP Apps, the Stainless acquisition — and what’s emerging is a fully vertically integrated AI company: chips at the bottom, frontier models in the middle, agent infrastructure and enterprise rails at the top.
The risk in this trajectory isn’t demand — the run-rate number speaks for itself — it’s the operational complexity of being all of those things at once. Each layer (chips, models, agents, enterprise distribution) has different metabolism and different talent pools, and the “ad-free” commitment is partly a values statement and partly a strategic narrowing: enterprise margins or bust. As long as the SWE-Bench Pro deltas keep landing and the SAP-class deals keep stacking, that bet looks like the right one. The question that’ll define the back half of 2026 is whether anyone can keep up.