AI Bulletin · week of May 25

Recap of the week of May 25–29, 2026 in AI. Six news items and a framework on the economics of AI spending that circulated that week.

1. Anthropic ships Opus 4.8, closes a $65 billion Series H at a $965 billion valuation and introduces Dynamic Workflows in Claude Code

Anthropic has released Claude Opus 4.8 with benchmark improvements, adjustable effort controls, dynamic workflows in Claude Code and a cheaper fast mode. In his notes, Simon Willison describes the model as a modest but tangible improvement, with more uncertainty flagging and honesty and a new kind of mid-conversation system message.

The company also announced the close of a $65 billion Series H round at a $965 billion post-money valuation, citing a $47 billion run-rate revenue. CNBC places Anthropic close to a $1 trillion valuation and as the most valuable AI company in Silicon Valley.

In parallel, Anthropic has introduced Dynamic Workflows in Claude Code as a research preview in the CLI, desktop and VS Code for the Max, Team and Enterprise plans, plus the API, Amazon Bedrock, Google Vertex AI and Microsoft Foundry. The feature runs orchestration scripts that spawn tens to hundreds of parallel subagents, with high token usage. Anthropic cites the case of Jarred Sumner, who rewrote Bun from Zig to Rust with a 99.8% test success rate and 750,000 lines in eleven days.

Sources: Anthropic — “Claude Opus 4.8” and “Series H”; Anthropic — “Introducing dynamic workflows in Claude Code”; Simon Willison — “Claude Opus 4.8”, 28 May 2026; CNBC — “Anthropic nears $1 trillion valuation in latest round”, 28 May 2026.

The bulletin mascot at a newsroom desk with ticker tape spooling out of a machine, a coffee mug and an open notebook, reviewing a freshly printed sheet — Anthropic: Opus 4.8, a $65 billion Series H round at a $965 billion valuation and Dynamic Workflows in Claude Code.

2. The Anthropic–SpaceX compute lease: conflicting accounts of how long it runs

The compute deal between Anthropic and SpaceX that this series covered last week as a three-year commitment has produced conflicting accounts of its actual duration. According to TechCrunch, several reports describe the contract as a 180-day lease with the option of mutual cancellation on 90 days’ notice, a characterisation that clashes with the three-year description in SpaceX’s IPO prospectus and with the way Elon Musk has presented it in public.

In the same infrastructure block, a report notes that SpaceX has almost finished writing version 1.0 of an in-house AI training stack in C, with pipeline parallelism exact-mapped to 220,000 GB300s connected via 800G NICs, and that its next target is an inference stack also written in C.

Sources: TechCrunch — “How long is Anthropic’s lease with SpaceX? Opinions vary”, 28 May 2026; report via TLDR AI, 29 May 2026 — “SpaceX has almost finished writing v1.0 of an in-house AI training stack in C”.

3. Coding agents: Microsoft prepares an in-house model and retreats from Claude Code, xAI launches Grok Build and Cursor publishes usage data

Microsoft is developing a new in-house coding model, according to a report carried by Sherwood News, in an attempt to get back into competition in AI for code. In parallel, The Next Web has reported that the company is cancelling most direct Claude Code licences in one of its groups and has asked its developers to move to GitHub Copilot CLI by June 30, a continuation of the internal pivot this series covered last week.

xAI has launched Grok Build, a coding agent with a CLI in beta for SuperGrok and X Premium Plus subscribers, with a plan mode with review, a headless mode and subagents.

Cursor has published its developer habits report, according to which models use more and more context to understand codebases, and a lower cost of input and cache-read tokens is associated with higher productivity and with greater survival of the diffs they propose.

Sources: Sherwood News — “Report: Microsoft tries to get back in the AI coding game with new model”, 28 May 2026; The Next Web — “Microsoft’s quiet Claude Code retreat and the real cost of enterprise AI”; report via TLDR AI, 26 May 2026 — “Introducing Grok Build”; Cursor — “The Cursor Developer Habits Report”.

4. Regulation and governance: a study on EU compliance, OpenAI’s frontier framework and a papal letter on AI

A study by the nonprofit research foundation Aithos concludes that every major model tested violates EU AI and data protection regulations to varying degrees, with issues around consent, profiling, manipulation and the handling of sensitive information. According to the study, some models failed compliance checks in up to 93% of scenarios.

OpenAI has published a frontier governance framework that addresses aligning technical safety and information security with emerging regulation, with sections on risk management, model reporting, incident response and oversight.

Pope Leo XIV has published a document on AI that compares its threat to the biblical Tower of Babel and warns that the technology may reduce people to cogs in a machine and concentrate power in a few private actors. Simon Willison has published his notes on the encyclical.

Sources: Computerworld — “Study finds all major AI models violate EU regulations”; report via TLDR AI, 29 May 2026 — “OpenAI published a Frontier Governance Framework”; Simon Willison — “Notes on Pope Leo XIV’s encyclical on AI”, 25 May 2026.

The bulletin mascot with a classic microphone in front of an empty lectern, next to a balance scale and a stack of blank labelled folders on the oak desk — The week's governance debate: model compliance in the EU, OpenAI's frontier framework and Pope Leo XIV's letter on AI.

5. Consolidation of enterprise agent governance: Snowflake buys Natoma, Asana buys StackAI and Cohere buys Reliant AI

The control and governance layer for enterprise agents concentrated several deals this week. Snowflake announced its intent to acquire Natoma, a centralised MCP gateway that enforces identity, policy and audit at the tool-call level, with the goal of having its Cortex Agents connect to corporate applications through a single governed interface.

Asana has acquired StackAI, a no-code platform for building AI agents that execute workflows across systems such as Salesforce, Oracle, AWS and DocuSign. Cohere has bought Reliant AI to strengthen its sovereign enterprise AI in healthcare and life sciences, adding biomedical datasets and a research team.

Along the same lines, Anthropic has added 28 security and compliance integrations for Claude, covering identity, DLP, SIEM and AI governance, and Salesforce has launched, in Developer Preview, the Data 360 MCP server, a single interface for any MCP client to operate on its data. Workday and Google Cloud have expanded their partnership to bring AI agents closer to HR and finance workflows.

Sources: Snowflake — “Snowflake to acquire Natoma to bring governed agentic access to the enterprise”; SiliconANGLE — “Asana acquires StackAI to run AI agent workflows across enterprise systems”, 28 May 2026; Cohere — “Cohere acquires Reliant AI to expand sovereign enterprise AI”; Help Net Security — “Anthropic adds 28 security and compliance integrations for Claude”, 25 May 2026; Salesforce — “Introducing the Data 360 MCP Server”.

6. Funding and infrastructure: Cognition at $26 billion, OpenRouter at $1.3 billion, Glean past $300 million, and Mistral and ByteDance weigh their own chips

Cognition has raised more than $1 billion at a $26 billion valuation to expand Devin, its coding agent, and cites Mercedes-Benz and Itaú among its clients. OpenRouter has more than doubled its valuation to $1.3 billion with a $113 million Series B led by CapitalG, with more than 400 models available and 100 trillion tokens processed per month.

Glean has said it has crossed $300 million in annualised revenue, triple its figure fifteen months ago, and is increasingly positioning itself as a way to reduce model consumption costs by giving AI systems better enterprise context and limiting unnecessary token usage.

In the infrastructure block, Mistral CEO Arthur Mensch has said the company is exploring designing its own chips to control infrastructure and costs as it expands its data centres in the EU. ByteDance, for its part, has approached several partners to design its own AI infrastructure chip.

Sources: Cognition — “Series D”; TechCrunch — “OpenRouter more than doubles valuation to $1.3B in a year”, 26 May 2026; TechCrunch — “Glean’s revenue crosses $300 million as AI cost reduction becomes a major selling point”, 28 May 2026; CNBC — “Mistral to explore designing own chips, CEO says”, 28 May 2026; PC Gamer — “ByteDance has had enough of waiting months for processors, so it’s going to make them itself”.

The bulletin mascot in front of a ticker board with rising and falling lines, holding a notebook next to a tray with stacks of coins and a silicon chip on the desk — Funding and infrastructure: rounds from Cognition, OpenRouter and Glean, and Mistral's and ByteDance's plans for their own chips.

7. Framework of the week: from “are they using it?” to “what did each token buy?”

Several pieces that circulated this week converge on the economics of AI spending and on how its return is being measured.

A piece distributed by TLDR Founders on the “token budget wars” argues that the relevant question shifts from usage to tracking tokens per outcome: cost per resolved ticket, per processed claim or per unit of work completed.
An analysis published by SiliconANGLE notes that FinOps is moving from cloud optimisation to executive-level planning, with the focus on measuring agent return, model consumption and long-term AI operating expenses.
A piece on Dave Friedman’s blog argues that the “AI pullback” people have talked about has not happened: pilots die while production usage and token consumption rise, and the real constraint is measurement.
An article in CIO states that companies are spending billions to train their workforce in AI, and that much of that spend will fail unless workflows and incentives are redesigned.
A DataGrail report carried by VentureBeat finds that, of 2,400 vendors analysed, 63.6% of those advertising AI features did not disclose all the third-party AI subprocessors they use.

Sources: report via TLDR Founders, 29 May 2026 — “Token Budget Wars”; SiliconANGLE — “FinOps is becoming a boardroom conversation for AI spending”, 28 May 2026; Dave Friedman — “The AI pullback that didn’t happen”; CIO — “Companies are spending billions to train workers for AI. Most of it will fail”; VentureBeat — “DataGrail report finds your vendor may be sending data to AI models you never approved”.

Closing

The six items above cover the release of Claude Opus 4.8 alongside Anthropic’s largest funding round to date, the conflicting accounts of how long the SpaceX compute lease runs, the moves in coding agents from Microsoft, xAI and Cursor, a study on model compliance in the EU together with OpenAI’s governance framework and a letter from Pope Leo XIV, the consolidation of the enterprise agent governance layer through several acquisitions, and a batch of funding rounds and in-house chip plans. The framework of the week gathers five independent readings on how AI spending and its return are being measured. None of the items published include verified productivity data measured in specific companies.