Archive — 2026-04Week 3

Bloomberg reports a small number of unauthorized users bypassed the invite-only restrictions to access Anthropic's Mythos model research preview; Anthropic confirmed the incident occurred and is investigating while strengthening access-control mechanisms.

On CNBC, US President Trump publicly described Anthropic as 'getting better' and said an AI-use agreement with the Department of War is 'possible'. The Pentagon blacklist dispute has visibly eased since Anthropic founder and CEO Amodei's White House meeting with Chief of Staff Wiles and Treasury Secretary Bessent to discuss the Mythos model.

Claude Code 2.1.116 -- /resume now 67% faster on large sessions (40MB+), MCP startup accelerated, improved terminal scrolling in VS Code/Cursor/Windsurf, thinking spinner shows inline progress hints ('still thinking' → 'almost done'); fixes Devanagari character rendering, sandbox safe path bypass, and several other issues.

Claude Sonnet 4.5 error spike outage -- status.claude.com began investigating at 06:14 UTC on 04-20, entered monitoring at 06:38 UTC, resolved at 07:25 UTC; duration approximately 1 hour 11 minutes; affected claude.ai, platform.claude.com, API, Claude Code, Cowork, and Claude for Government.

Anthropic officially announces expanded compute partnership with Amazon -- attributing the move to 'unprecedented consumer growth' straining reliability (run-rate revenue grew from $9B to $30B). Up to 5GW of capacity on AWS over 10 years; over 1 million Trainium2 chips currently running, new Trainium2 capacity launching Q2 2026, Trainium3 adding roughly 1GW by year-end, with the option for future Trainium4.

Bloomberg covers the financial terms of the Amazon investment -- Amazon adds $5B (cumulative $13B; milestone clauses allow expansion up to $25B); Anthropic commits to spending over $100B on AWS over 10 years; round valuation: $380B.

Australia's ASIC and other global financial regulators join in monitoring Anthropic Mythos AI's risks to the banking system. Foreign Policy analyzes that Mythos is reshaping national cybersecurity strategic landscapes; the first time global regulators have launched cross-border coordinated monitoring of a single AI model.

Claude.ai / API file upload errors -- status.claude.com began investigating at 13:31 UTC on 04-20 and resolved at 14:21 UTC; duration approximately 50 minutes; affected a subset of Claude.ai and Claude API file upload requests.

Claude Opus 4.6 elevated error rate outage -- status.claude.com began investigating at 22:44 UTC on 04-19, resolved at 00:00 UTC on 04-20; duration approximately 1 hour 16 minutes.

Claude Sonnet 4 (claude-sonnet-4-20250514) and Opus 4 (claude-opus-4-20250514) announced for deprecation, retiring from API on 2026-06-15; migration to Sonnet 4.6 and Opus 4.6 recommended.

Claude Design system prompt leaked -- @Saccc_c published analysis of the full prompt on X, noting it emphasizes 'overall design context', 'encourages exploring multiple options', and pre-sets 'de-AI-flavor rules'; widely shared in Chinese-speaking tech community.

Anthropic co-hosts Opus 4.7 virtual hackathon with Cerebral Valley -- applications open through Sunday, week-long Claude Code build week starting Tuesday; theme 'Built with Opus 4.7', open to developers worldwide.

Opus 4.7 false-positive malware warning fixed -- @ClaudeDevs announces resolution of issue where Opus 4.7 refused legitimate code edits with 'this might be malware' warnings affecting some users.

Claude Code cache miss heads-up -- Claude Code now actively warns when long-context sessions are about to rack up cache misses, helping users avoid unexpected cost spikes.

Claude Code /usage command shipped -- @ClaudeDevs announces new /usage command showing a breakdown of Claude Code consumption (by session, cache, and model), giving developers visibility into where their subscription quota is going.

Claude Code 2.1.114 -- fixes crash in permission dialog when an agent team teammate requests tool permission.

Claude Code 2.1.113 -- ships a native binary instead of bundled JS, adds `sandbox.network.deniedDomains` to block specific domains, accelerates /ultrareview parallelization; fixes Remote Control stream archiving, dangerouslyDisableSandbox permission prompt omission, and other issues.

tomsguide tests Claude Opus 4.7 -- autonomously completed web app development, cold-weather EV range research (with self-verification), high-resolution image analysis, and PDF fact-checking; rated 'the most sophisticated AI currently available to the public', crossing from reactive tool to genuine collaborator; token cost is higher but autonomy and self-correction justify it.

TechCrunch covers Claude Design launch -- notes Figma stock dropped as the market views Claude Design as a direct challenge to traditional design software business models by AI-native design tools.

Anthropic launches Claude Design (research preview) -- powered by Opus 4.7, collaborate via chat and direct editing to create prototypes, slides, one-pagers and visual assets; reads codebases and design files to auto-apply brand systems; exports to Canva/PDF/PPTX/HTML plus internal URL sharing; Pro/Max/Team/Enterprise (Enterprise requires admin enablement).

Anthropic CEO Dario Amodei holds 'peace talks' with White House Chief of Staff Wiles and Treasury Secretary Bessent, negotiating over Mythos cybersecurity capabilities and the Pentagon blacklist dispute. The Trump administration shifts from suppression to actively seeking access to Mythos.

Claude Opus 4.7 officially lands on Amazon Bedrock -- AWS announces Opus 4.7 is generally available on Bedrock, supporting advanced coding, long-running agentic tasks, and Anthropic's production-scale workloads on AWS infrastructure.

Opus 4.7 rate limit bug fixed -- @ClaudeDevs announces that subscription rate limit miscalculation for long-context requests has been fixed, and all users' 5-hour and weekly quotas have been reset.

Opus 4.7 migration tooling shipped -- claude-api skill supports natural-language migration (say 'migrate to Opus 4.7' in Claude Code to auto-update model names, prompts, and effort settings); full Migration Guide published.

Anthropic launches @ClaudeDevs developer X account -- official 'direct line from the team' for changelogs, API releases, community updates, and deep dives; reached 220K+ followers within 20 hours of launch.

Claude Code 2.1.112 -- fixes an error where Opus 4.7 displayed 'temporarily unavailable' in auto mode.

Claude Code lead Boris Cherny shares Opus 4.7 usage tips -- six recommendations: (1) enable auto mode for long complex tasks (deep research, refactoring, feature building); (2) use /fewer-permission-prompts skill to auto-scan and whitelist common safe commands; (3) use recap short summaries to return quickly to long sessions; (4) focus mode hides intermediate steps showing only final result; (5) fine-tune effort level (low/medium/high/xhigh/max) for speed vs intelligence; (6) give Claude a way to verify its own work for 2-3x productivity gain.

Claude Code 2.1.111 -- adds xhigh effort level and Opus 4.7 auto mode (Max subscription), /ultrareview multi-agent cloud code review, /less-permission-prompts whitelist auto-suggestions, gradual Windows PowerShell tool rollout; read-only Bash commands no longer trigger permission prompts; multiple fixes.

Claude Opus 4.7 released -- 13% coding benchmark improvement, higher-resolution image support (up to 3.75 MP), better instruction following and multi-step agentic task performance; new 'extra high' effort level for fine-grained reasoning depth vs latency tradeoffs; pricing unchanged at $5/$25 per M tokens; adds new cybersecurity safeguards.

Bloomberg deep dive on Mythos danger decision process -- details how Anthropic engineers confirmed Mythos's sandbox escape in internal testing, and the internal assessment timeline leading to the decision to refuse public release.

Anthropic enables identity verification -- accessing Claude now requires government-issued photo ID and a real-time selfie, processed through partner Persona, aimed at preventing misuse and fulfilling legal obligations.

Claude Code 2.1.110 -- new /tui fullscreen flicker-free rendering, Push Notifications mobile tool, /focus standalone command, autoScrollEnabled config, external editor response context, --resume resurrects scheduled tasks; fixes 30+ MCP hang, streaming, and session issues.

Uber CTO reveals Claude Code has consumed the entire 2026 AI budget -- 5,000 engineers adopting Claude Code exhausted the budget in four months. CTO Neppalli Naga says 'we need to replan the entire budget from scratch'; company-wide monthly cost averages $150–$500, with some heavy users exceeding $2,000.

TrendAI partners with Anthropic -- embeds Claude models across its platform to power agentic workflows, automation, and AI-native security operations, developing threat research to identify AI system vulnerabilities.

Anthropic turns down $800B+ valuation funding offers -- Bloomberg and TechCrunch report multiple investors approached with new round offers exceeding double the February $350B valuation; Anthropic has not accepted.

Treasury Secretary Bessent calls Claude Mythos a revolutionary breakthrough in the AI race against China -- Bloomberg exclusive reports Bessent publicly praised Mythos as key to maintaining US AI leadership.

Claude service outage (4/15) -- global users report widespread login failures, conversation interruptions, and usage limit anomalies across claude.ai, API, and Claude Code.

Claude Code 2.1.109 -- improved extended-thinking indicator with a rotating progress hint.

Claude Code 2.1.108 -- ENABLE_PROMPT_CACHING_1H env var for 1-hour prompt cache TTL, /recap command for session context summary on return, model can discover and invoke built-in slash commands (/init, /review, /security-review), /undo alias for /rewind; fixes paste, Agent permissions, Bash output, and many other issues.