The Weekly AI Digest
Week of 25–31 May 2026
Top 5 AI Stories This Week
Anthropic Closes $30B Round at $900B Valuation, Surpassing OpenAI
Anthropic closed a round topping $30 billion at a valuation above $900 billion, vaulting past OpenAI to become the world's most valuable private AI company. Sequoia, Dragoneer, Altimeter and Greenoaks each committed roughly $2 billion as co-leads. Annualised revenue is now $30 billion (from $9 billion at end of 2025), with Q2 projected at $10.9 billion as the first profitable quarter. An IPO is reported as expected by October.
Why it matters: A 2.4x valuation step-up in 14 weeks from the February Series G. For enterprise buyers, the round confirms Claude's long-term commercial viability at a scale that reduces the vendor risk questions that periodically surface.
Microsoft Cancels Internal Claude Code Licenses
Microsoft is cancelling Claude Code licenses across its Experiences and Devices division (Windows, Microsoft 365, Outlook, Teams, Surface engineering teams) with a hard cutoff of June 30, redirecting engineers to GitHub Copilot CLI. Uber CTO Praveen Neppalli Naga separately disclosed that the company burned through its 2026 AI coding budget in four months. GitHub announced that all Copilot plans shift to usage-based billing via GitHub AI Credits from June 1.
Why it matters: Flat seat licensing made token consumption invisible during pilots, and true cost only became visible at scale. Token caps and consumption budgets now need to sit alongside licensing terms in any agentic coding agreement.
DeepSeek Makes 75% V4-Pro Price Cut Permanent
DeepSeek announced on May 22 that the 75% promotional discount on V4-Pro becomes the permanent list price. V4-Pro pricing drops from $1.74/$3.48 to $0.435/$0.87 per million input/output tokens, with cache hits at $0.003625. DeepSeek attributes the cut to V4-Pro's architecture, citing roughly one-quarter of the single-token compute and one-tenth of the memory footprint of its predecessor.
Why it matters: Frontier-tier open weights now sit at roughly 11x cheaper than Claude Opus 4.7 on input. For high-volume workloads (codebases, agent loops, batch processing), self-hosting or hybrid routing becomes a real procurement consideration.
OpenAI Publishes Frontier Governance Framework
OpenAI published its Frontier Governance Framework on May 28, mapping how its internal safety practices align with California's SB 53 and the EU AI Act's Code of Practice for General Purpose AI. The framework covers four risk areas (cyber offense, CBRN, manipulation, loss of control) plus model reporting, incident response and external expert input. It sits alongside the existing Preparedness Framework rather than replacing it.
Why it matters: Frontier model governance is moving from voluntary commitment to public compliance disclosure aligned with specific statutes. The framework gives procurement and risk teams a structured document to cross-reference against EU and California requirements without bespoke legal interpretation per vendor.
OpenAI and Thrive Holdings Ship Self-Improving Tax AI on Codex
OpenAI published a case study on May 27 covering a tax preparation system built with Thrive Holdings and deployed through Crete Professionals Alliance. The system processed 7,000 returns in pilot, with draft accuracy rising from around 25% to 97%, preparation time down by one-third, and throughput up 50%. Practitioner corrections become structured evaluation criteria that Codex uses to refine the system through bounded, validated engineering tasks.
Why it matters: One of the more concrete examples of production AI using human feedback as a continuous training signal rather than waiting for model upgrades. Vertical applications with high-volume structured workflows and human-in-the-loop validation (tax, claims, document review) are where this pattern translates most directly.