🧭 Anthropic Opens Singapore Office and Signs Government MOU — APAC's Second-Largest Claude Market Gets a Permanent Home

Anthropic has opened its Singapore office and signed a Memorandum of Understanding with Singapore's Ministry of Digital Development and Information. Anthropic co-founder Ben Mann signed the agreement alongside Permanent Secretary Kai Fong Chng and Minister Josephine Teo. The MOU covers cooperation on AI safety and security evaluations, joint research on AI adoption impact, and frameworks for responsible AI deployment across Singapore's public and private sectors. Singapore ranks second among 121 countries in Claude.ai usage per capita — a striking adoption figure that made a dedicated office a natural next step.

Who is already using Claude in Singapore

• GovTech Singapore: 1,500+ engineers on Claude Code — the largest government deployment of Claude Code globally by headcount
• DBS Bank: Southeast Asia's largest bank, deploying Claude across engineering and operations
• Grab, Sea, Supabase: Three of the region's highest-profile tech companies among early adopters

What the MOU means in practice

• Joint AI safety evaluations — Singapore's AI Verify Foundation and Anthropic will co-develop testing frameworks aligned with international standards
• Research on AI adoption impact — academic partnerships with NUS and NTU to study Claude's economic and social effects in the region
• Responsible adoption guidance — model governance playbooks for Singapore's public sector agencies

🧭 Anthropic Prepares Claude for the 2026 US Midterms — Real-Time Voting Data, 99% Evenhandedness, State-Actor Classifiers

Anthropic has published a detailed account of how it is preparing Claude for the 2026 US midterm elections. The report combines usage data, model evaluation scores, third-party audits, and enforcement actions. Key new capability: Claude now surfaces real-time election night results in-chat, and from late October will display state-specific official voting information — polling locations, hours, and ID requirements — directly in responses, powered by a partnership with non-partisan nonprofit Democracy Works (TurboVote). Election-related topics appear in only 0.51% of US conversations; requests asking Claude whom to vote for appear in 0.01%.

How Anthropic is testing political balance

• Internal eval — 99% political evenhandedness score on Claude's latest models across a battery of contested topics
• Three independent auditors — The Future of Free Speech, Foundation for American Innovation, and Collective Intelligence Project — reviewed Claude's handling of controversial queries and flagged two weaknesses: Claude sometimes adds unsolicited counterarguments and doesn't always hedge symmetrically across the political spectrum
• Blind spot acknowledged: The evaluations found Claude occasionally adds unsolicited counterarguments to partisan requests — a form of "balance injection" the company says it is working to make optional rather than default

Enforcement actions taken

• New multi-step campaign classifiers deployed to detect coordinated influence operations using Claude-generated content
• Accounts running fake news sites producing midterm coverage with Russian, Iranian, and North Korean narratives were discovered and banned
• Financially motivated election content schemes (ad-revenue farms using fabricated election results) also detected and removed

🧭 Anthropic Replaced Git With an In-House VCS Called Scion — 30,000 Daily Merges, 200× Smaller Infrastructure

Anthropic's engineering blog has published a detailed account of how the company's monorepo outgrew Git and how a team of seven engineers built a replacement — called Scion — in five months. The trigger: merge volume exploded from 2,500 per day in May 2026 to over 30,000 by September, driven largely by Claude Code automating developer workflows. The monorepo's git history alone consumed nearly 100 GB, and CI was performing ~25 million code checkouts daily using hundreds of internal git mirrors just to keep up. The team also built Antique, an in-house code-review tool, in parallel.

Scion's core design principles

• Always-online clients: Unlike Git's offline-first design, Scion assumes clients (agents and CI jobs) are permanently connected — eliminating the full-history local clone entirely
• Client-server asymmetry: Complex operations (commits, merges, blame) run server-side; clients maintain only minimal local caches
• Horizontal scalability: No single copy of the repository exists anywhere — data is sharded across servers
• Optional working copies: Many operations (most CI steps) query the server directly without materialising files to disk at all

Results after migration

• Infrastructure footprint: 200× smaller than the previous git-mirroring setup
• CI checkout times: two orders of magnitude faster (seconds, down from minutes)
• 25% CPU capacity recovered for CI alone — effectively a free capacity upgrade

# Git shallow clone as a stopgap (reduces checkout data significantly)
git clone --depth=1 --filter=blob:none --sparse https://your-repo.git
git sparse-checkout init --cone
git sparse-checkout set path/to/relevant/subdir