Archive — 2026-05Week 3

Boris Cherny, creator and head of Claude Code, previewed on X that the next version of Claude Code will add a `/usage` command breaking down which Skills, Agents, MCPs, and Plugins are consuming tokens (his example shows an MCP server "testmcp" accounting for 11% of usage, per-skill and per-subagent usage percentages, and day/week views); the feature lands on the CLI first, with the Desktop app to follow.

Bloomberg (reporters Matt Day and Robin Ajello), citing The Information, reports that Anthropic is in early talks to rent Microsoft's Maia AI server chips to expand its compute capacity and meet growing demand for Claude services. The Maia chips would give Anthropic an infrastructure path for running Claude models beyond Nvidia; the move echoes an industry trend of large tech companies — Microsoft, Google, and Amazon — building their own chips to reduce dependence on Nvidia. The talks are still at an early stage and involve renting rather than buying outright.

Anthropic launched the Claude Compliance API alongside 28 security and compliance tool integrations. The Compliance API exposes two data streams: (1) programmatic access to Claude Enterprise conversation content (chats, files, projects) for security monitoring; (2) activity event data across Claude Enterprise and Platform (logins, admin actions, configuration changes) for unified governance visibility. The 28 launch partners span DLP, SASE, data security, SIEM/security operations, identity management, eDiscovery, AI security posture management, and observability. Named partners: Cloudflare, Cribl, CrowdStrike, Cyera, Datadog, Forcepoint, Fortinet, Geordie AI, IBM Guardium, Microsoft Purview, Mimecast, Netskope, Okta, Palo Alto Networks, Proofpoint, Relativity, ReliaQuest, Rubrik, SailPoint, Smarsh, Snyk, Sumo Logic, Tenable, Theta Lake, Trellix, Varonis, Wiz, and Zscaler. Once organizations connect their Claude instance to a supported platform, data flows directly into existing dashboards and alerting systems.

Anthropic's blog showcased partner results using Claude Opus in cybersecurity workflows. Launch partners include Wiz, Palo Alto Networks, CrowdStrike, Accenture, TrendMicro, Deloitte, and PwC, with BCG, Infosys, and SentinelOne still in development. Specific outcomes: Wiz Red Agent performs continuous pentesting across 150,000+ production assets per week, surfacing thousands of validated high/critical findings weekly; Palo Alto Networks Unit 42 delivers vulnerability mapping and AI attack-path analysis, completing a year's worth of pentesting effort in under three weeks; Accenture Cyber.AI lifted test coverage across 1,600 applications and 500,000+ APIs from ~10% to 80%, cutting scan time from 3–5 days to under one hour; TrendMicro uses Opus to identify exposures and enable virtual patching, blocking up to 96 days before vendors' official patches ship; PwC pitches moving enterprises from sandbox to production AI deployment "in weeks rather than quarters," with audit evidence and defined guardrails. The post argues Opus's code reasoning, real-world risk assessment, and sustained agentic workflows underpin these offensive-testing, remediation-acceleration, and governed production deployment use cases.

Fortune reports Anthropic brought its Code with Claude developer conference to Europe for the first time (a two-day event held in London on 5/19), coinciding with AI-assisted coding — and the anxieties around it — going mainstream. At the event Anthropic rolled out two enterprise-focused features for Claude Agents — sandboxes that let companies run agents on their own infrastructure, and MCP tunnels that let agents reach internal systems without going over the public internet — both addressing enterprise concerns about the security and control of AI systems. Claude Code head Boris Cherny compared manual coding to analog photography: "There's always room for that." Co-founder Jack Clark lectured at Oxford on AI's philosophical implications, including the "non-zero chance" of catastrophic outcomes; head of engineering Fiona Fung advised companies to start automating with their most problematic workflows. The report cites industry projections of $2.59 trillion in worldwide AI spending in 2026 (up 47% year-over-year) and a forecast 110% rise in model-consumption spending, and notes Anthropic will pay SpaceX $15 billion a year for compute through May 2029.

MIT Technology Review reports on Anthropic's Code with Claude developer conference held in London on 5/19 (a two-day event also run in San Francisco and Tokyo), portraying how AI-assisted coding is reshaping software development. The venue was packed, and when the host asked who had shipped a pull request written entirely by Claude, nearly half the audience raised their hands; Anthropic says "most software at Anthropic is now written by Claude." The report also describes Claude Code's "dreaming" mechanism — agents take their own notes while working on tasks, and the dreaming system consolidates those notes across tasks to help agents learn from past errors and patterns. Claude engineering lead Katelyn Lesse said Claude is "probably as good as a midlevel engineer at writing code," though expert engineers are still needed for system design and complex troubleshooting. Outside the conference, developers voiced concerns: the review burden of AI-generated code, doubts about real productivity gains, skill atrophy from over-reliance, and the security risk of unsafe generated code.

CFO.com reports that Anthropic's Claude is rapidly penetrating corporate finance workflows — Claude has been embedded into forecasting, reporting, month-end close, and audit preparation processes at major financial institutions and accounting firms. The article surveys recent moves: PwC will train 30,000 U.S. professionals on Claude, and KPMG has integrated Claude into its Digital Gateway for tax and private equity clients; JPMorganChase CEO Jamie Dimon has tested Claude for market analysis, Goldman Sachs CIO Marco Argenti has discussed AI's operational impact, and Citi, AIG, and Visa are also among adopters. The piece also surfaces CFO-side reservations: Spindrift CFO Karim Sadik-Khan questions Anthropic's enterprise support and pricing transparency, while Nate Saperia of Saperia Consulting notes that the financial services agents Anthropic launched on May 5 mainly target institutional workflows rather than corporate finance (FP&A) itself.

Claude Code 2.1.146 — Renamed `/simplify` to `/code-review` with an optional effort level (e.g. `/code-review high`); auto mode no longer suppresses `AskUserQuestion` when the user or a skill explicitly relies on it. Fixes: the Windows PowerShell tool failing with "command line is invalid" when `pwsh` is installed via winget or the Microsoft Store (regression in v2.1.124); MCP `resources/list`, `resources/templates/list`, and `prompts/list` dropping items past page 1 on paginating servers; full-screen strobing in attached background sessions on Windows Terminal while streaming; the auto-updater status line not showing the current version when an update fails; on Windows, removing a background-job worktree following NTFS junctions into the main repo; `/background` refusing sessions whose only typed input was a skill or custom slash command; backgrounded sessions re-prompting for tool permissions already granted with "don't ask again"; the `/theme` color editor and "New custom theme" dialogs not responding to Esc; an uncaught exception at the end of streaming sessions run via the Agent SDK; the `forceLoginOrgUUID` and `forceLoginMethod` managed-settings policies not being enforced against third-party-provider and API-key sessions; GNOME Terminal right-click and middle-click paste not inserting text; `CLAUDE_CODE_SUBAGENT_MODEL` not being forwarded to child processes in multi-agent sessions. Improvements: auto-updater reliability (native version checks and downloads now retry transient network failures instead of failing immediately); diff rendering performance for large file edits.

The Register's Jessica Lyons reported on Wyze Labs security researcher Aonan Guan's disclosure of two Claude Code sandbox bypass vulnerabilities — one data-exfiltration bug was patched in Claude Code 2.1.88 (3/31) after roughly 5.5 months of exposure from sandbox GA through v2.1.90. The exploit: SOCKS5 hostname null-byte injection bypasses the allowlist filter; combined with prompt injection, it tricks Claude into reading hidden instructions and running attacker code inside the sandbox, exfiltrating cloud credentials, GitHub tokens, metadata, and internal APIs. Anthropic says it found and patched the flaw before receiving Guan's 4/3 report, with the fix shipped as a public commit in the sandbox-runtime repo, but issued no CVE or security advisory for Claude Code itself (the related CVE-2025-66479 was assigned only to the underlying sandbox-runtime library and patched in v0.0.16). This is the second major Claude Code vulnerability Anthropic has patched without public disclosure in five months. Guan: "Shipping a sandbox with a hole is worse than not shipping one. The user with no sandbox knows they have no boundary. The user with a broken sandbox thinks they do."

Claude Opus 4.7 elevated error rate — status.claude.com began investigating at 05-20 14:43 UTC, deployed a fix and moved to monitoring at 15:56 UTC, and resolved at 16:04 UTC, lasting roughly 1 hour 21 minutes; impact spanned claude.ai, Claude API (api.anthropic.com), Claude Code, and Claude Cowork; Anthropic did not disclose the root cause.

Claude Haiku 4.5 elevated error rate — status.claude.com began investigating at 05-20 08:14 UTC, deployed a fix and moved to monitoring at 08:31 UTC, and resolved at 08:49 UTC, lasting roughly 35 minutes. Affected components: claude.ai, Claude Console (platform.claude.com), Claude API (api.anthropic.com), Claude Code, and Claude Cowork. Anthropic did not disclose root cause.

A Claude blog post titled "Using Claude Code: The unreasonable effectiveness of HTML" by Anthropic Member of Technical Staff Thariq Shihipar (personal opinion) argues HTML beats Markdown as a Claude Code output format. HTML supports richer visualizations, color, and diagrams while remaining more readable for documents over ~100 lines. On information density, HTML accommodates tables, CSS styling, SVG illustrations, JavaScript interaction, and spatial layout, avoiding Markdown workarounds like ASCII diagrams. On navigation, tabs, links, and responsive design reinforce structural cues, making long specs (typically ignored as Markdown) more digestible. On shareability, link-distributable HTML circulates more easily than Markdown attachments, raising the chance stakeholders actually read it. On two-way interaction, HTML supports sliders, toggles, and tunable parameters; users can experiment within the document and feed results back into prompts or version control. Recommended workflow: generate exploratory HTML, expand chosen options with mockups and diagrams, produce an implementation plan, then hand off to a new session with full context for development. The technique is minimal — prompt Claude Code with "make an HTML file" or "make an HTML artifact," no dedicated skill required.

Anthropic's blog published an internal case from Travis Bryant, Head of US Mid-Market GTM, using Claude Cowork to manage 4,000 mid-market accounts. Workflows include automated daily meeting-room booking, customer call prep that pulls Salesforce and BigQuery data, Friday forecast reports for leadership, and an overnight account-propensity scoring system across all 4,000 accounts. Quantified results: micro-optimizations save roughly 90 minutes per day and 3 hours per week on forecast assembly; territory scoring that historically required "hundreds of hours across RevOps, FP&A, and marketing" completed in one night. Highlight: an interactive dashboard ranks scored accounts by propensity with per-dimension rationale and comparable case studies to direct prospecting activity.

Bristol Myers Squibb (BMS) announced a strategic agreement with Anthropic to deploy Claude Enterprise as a 'shared intelligence platform' across its global operations, connecting people, systems, and institutional knowledge for 30,000+ employees spanning research, clinical development, manufacturing, commercial, and corporate functions. Three primary implementation areas: engineering acceleration for software and AI development via Claude Code; workflow automation across R&D, drug development, manufacturing/quality, and commercial/medical affairs; and knowledge integration through secure connections to BMS's scientific and regulatory repositories. BMS Chief Digital & Technology Officer Greg Meyers said 'the real prize is the untapped value still trapped behind decades of data silos'; Anthropic Head of Life Sciences Eric Kauderer-Abrams noted Claude agents could 'generate a clinical study report from underlying trial data' and help medicines reach patients faster. No financial terms were disclosed.

Axios reports Anthropic has struck a large compute-purchase agreement with SpaceX — per SpaceX's IPO (S-1) filing submitted this Wednesday, Anthropic agreed to pay SpaceX $1.25 billion a month, roughly $15 billion a year, with the contract running through May 2029; payments are lower in May and June as a ramp-up period. The deal gives Anthropic compute capacity from SpaceX's Colossus 1 data center, expanding to Colossus 2, using Nvidia GB200. Either party may terminate with 90 days' notice. In the filing SpaceX said it expects to sign more similar compute-services contracts in the future while retaining the capacity needed to support its own AI models. The disclosure comes after Anthropic had told investors its annualized revenue topped $30 billion and that it expects to turn a profit for the first time next quarter, with massive compute-purchase spending a key variable in its cost structure.

CNBC's Arjun Kharpal interviews Boris Cherny, creator and head of Claude Code at Anthropic (previously a Principal Software Engineer at Meta) — Cherny says non-technical people are starting to code, that Claude Code has lowered the barrier to entry for software development, letting a wider range of employees take part in development work that previously required specialist skills and reshaping how enterprises allocate technical work. He cites his own team as an example: everyone on the Claude Code team codes — not just engineers, but the PM, designer, data scientist, finance person, and user researcher.

TechCrunch, reported by Lucas Ropek citing the Wall Street Journal — Anthropic told investors it expects to reach profitability for the first time in the upcoming (second) quarter. The company projects more than doubling revenue to around $10.9 billion in Q2 and delivering an operating profit for the first time. However, TechCrunch notes that because of the large compute costs Anthropic is scheduled to incur, it may not remain profitable throughout the year. The forecast was disclosed to investors during an ongoing funding round; Anthropic declined to comment further. The report notes the news coincides with market reports that OpenAI is preparing for an IPO as soon as September 2026, underscoring the competition between the two companies on revenue growth and capital-market timing; Anthropic's expanding market share reflects rising professional adoption of the Claude chatbot and its new offerings targeting small businesses and law firms.

Claude Code 2.1.145 — adds `claude agents --json` to list live Claude sessions as JSON for scripting (tmux-resurrect, status bars, session pickers); OTEL `claude_code.tool` spans gain `agent_id` and `parent_agent_id` attributes with fixed trace parenting for background subagent spans; status line JSON input now includes GitHub repo and PR information; `/plugin` Discover/Browse screens show a plugin's commands, agents, skills, hooks, and MCP/LSP servers before installation; `claude agents` terminal tab title shows the awaiting-input count; slash command and @-mention suggestion lists now support mouse hover and click in fullscreen; Stop and SubagentStop hook input adds `background_tasks` and `session_crons` fields; fixes a permission-prompt bypass where bare variable assignments to non-allowlisted environment variables in Bash were auto-approved; plus multiple bug fixes (spinner freezing after terminal resize, Windows PowerShell 5.1 cross-project resume, voice push-to-talk in agent view, task list ordering, PR badge updates, etc.).

Law & Crime reports oral arguments in Anthropic v. U.S. Department of War were heard at the U.S. Court of Appeals for the D.C. Circuit before a three-judge panel of Karen LeCraft Henderson (a George H.W. Bush appointee), Neomi Rao and Gregory Katsas (both Trump appointees). From the bench, Judge Henderson called the department's designation of Anthropic as a 'supply chain risk' a 'spectacular overreach', saying 'I don't see that the department has in any way supported its determination that there is a supply chain risk with Anthropic' and, 'for the life of me', she did not see 'any evidence of maliciousness' to justify the designation. Anthropic argued the action is 'textbook retaliation' built on false premises, contending Secretary of War Pete Hegseth had abandoned his original theory — that Anthropic held an 'operational veto' over Claude in deployed systems — and shifted to a new theory about future model limitations; the company also pointed to less restrictive alternatives available under federal procurement law. The case marks the first judicial interpretation of Section 4713 of the Supply Chain Security Act. Rao and Katsas appeared skeptical of some of Anthropic's arguments, with Henderson the only judge to show clear opposition to the Pentagon's actions.

TechCrunch follow-up by Rebecca Bellan and Lorenzo Franceschi-Bicchierai on Andrej Karpathy joining Anthropic (supplementing the 5/19 14:00 Axios entry) — Karpathy joins the pre-training team and reports to Anthropic's Pre-training Lead Nick Joseph (not directly to Dario Amodei); he will "start a team focused on using Claude to accelerate pre-training research." An Anthropic spokesperson confirmed pre-training is "responsible for the large-scale training runs that give Claude its core knowledge and capabilities." Karpathy wrote on X: "I've joined Anthropic. I think the next few years at the frontier of LLMs will be especially formative. I am very excited to join the team here and get back to R&D." The move came one day after the 5/18 Musk v. Altman trial concluded in Altman's favor and is read as a marker of intensifying OpenAI/Anthropic talent competition. Compensation, equity, and retention terms remain undisclosed.

Elevated errors on Claude Opus 4.7 — status.claude.com began investigating at 14:03 UTC on 05-19, implemented a fix and moved to monitoring at 15:19 UTC, and marked it resolved at 15:40 UTC, lasting about 1 hour 37 minutes; affected components spanned claude.ai, Claude Console (platform.claude.com), the Claude API (api.anthropic.com), Claude Code and Claude Cowork; Anthropic did not disclose a root cause.

Axios reports oral arguments in Anthropic v. U.S. Department of War at the D.C. Circuit Court of Appeals — the hearing surfaced a core contradiction in the Trump administration's posture: blacklisting Anthropic as a supply chain risk while actively evaluating its most powerful model Mythos for cyber defense. Government counsel Sharon Swingle argued Anthropic's "ideological" stance (refusing the "all lawful use" standard for military deployments) constitutes unreliability, and justified the Department of War's need to "act quickly" by invoking the "very real prospect of new red lines" and an "increasingly hostile posture" in negotiations. Anthropic counsel Kelly Dunbar countered that the government could simply "decline to do business with the company"; Congress requires agencies to consider less-intrusive alternatives before a supply-chain designation, making the permanent legal disbarment disproportionate. Trump-nominated Judge Gregory Katsas raised a technical concern beyond ideology: "the model is unpredictable" and "AI three months from now will be totally different," making discussion of current versions less meaningful. The D.C. Circuit acknowledged irreparable harm risk and agreed to expedite the case, with a possible ruling within weeks against the looming August Trump administration deadline.

9to5Mac reports Anthropic has rolled out two privacy and security features for Claude Managed Agents: MCP tunnels (limited research preview), which lets agents reach MCP servers inside private networks with no inbound firewall rules or public endpoints and end-to-end encryption; and Self-Hosted Sandboxes (public beta), which lets organisations execute agent tools within their own infrastructure while orchestration remains in Anthropic's cloud, with support for bring-your-own sandbox clients or partner providers Cloudflare, Daytona, Modal, and Vercel.

Axios reports OpenAI co-founder Andrej Karpathy is joining Anthropic — Karpathy announced Tuesday (2026-05-19) that he is joining Anthropic's pre-training team this week (the team responsible for the large-scale training runs that produce Claude's core knowledge and capabilities) and will help stand up a new team focused on using Claude itself to accelerate pre-training research. A founding member of OpenAI, Karpathy later served as Tesla's director of AI leading the Autopilot computer-vision team; after leaving Tesla in 2022 he became a prominent AI educator with large followings on YouTube and X, known for frontier-model technical explainers, and has described himself as being in a 'state of AI psychosis' since December 2025, deeply focused on model evaluation and development. Adding one of the field's most prominent AI researchers is seen as Anthropic doubling down on talent in its competition with OpenAI.

Axios Harris Poll 100 (14th annual corporate reputation survey) released — Anthropic ranks 15th overall in reputation among 100 companies; the partisan gap between Republicans and Democrats on Anthropic is only 1 point, the smallest among major AI firms (OpenAI's gap is 12 points), while other AI companies (OpenAI, Meta, Nvidia, X, TikTok) have all seen Republican favorability widen since 2024. Axios attributes Anthropic's relatively neutral cross-partisan reception to its principled stance — its earlier refusal to lift Claude safeguards against mass domestic surveillance or autonomous weapons development for the Trump administration angered the White House but unexpectedly produced cross-partisan equilibrium.

Elevated errors on Claude Opus 4.7 — status.claude.com began investigating at 08:41 UTC on 05-19 and marked it resolved at 10:08 UTC, lasting about 1 hour 27 minutes; affected components spanned claude.ai, Claude Console (platform.claude.com), the Claude API (api.anthropic.com), Claude Code and Claude Cowork; Anthropic did not disclose a root cause.

Elevated errors on Claude Haiku 4.5 — status.claude.com began investigating at 07:00 UTC on May 19, deployed a fix and entered monitoring at 07:56 UTC, and marked the incident resolved at 08:30 UTC, lasting roughly 1h 30m; affected services included claude.ai, claude.com, Claude Code, and the Claude API.

CNBC published its fourteenth annual Disruptor 50 list, with Anthropic taking the No. 1 spot for the first time and overtaking OpenAI (now No. 2) — CNBC noted this is the first time the ChatGPT maker has lost the top spot since the generative AI boom began. Anthropic's annualized revenue run rate has surpassed $30 billion, up from about $9 billion at the end of last year; the company said first-quarter revenue and usage grew roughly 80-fold on an annualized basis, far exceeding its internal planning estimate of 10x, and CEO Dario Amodei described the pace as 'just crazy ... too hard to handle.' CNBC attributed the top ranking not only to growth speed but to positioning — Anthropic's focus on building powerful AI systems that enterprises trust, centered on safety and constitutional AI; its Claude AI models, especially Claude Code, triggered a wave of selloffs among enterprise software stocks. Anthropic's in-progress new funding round could value it as high as $900 billion, potentially topping OpenAI. CNBC also released an interview with co-founder and president Daniela Amodei.

Claude Developer Platform official release note (2026-05-12): Fast mode (research preview) now supports the Claude Opus 4.7 model. By setting `speed: "fast"`, `model: "claude-opus-4-7"`, and the `fast-mode-2026-02-01` beta header in an API request, developers get significantly faster output token generation (up to roughly 2.5x output tokens/sec), priced at 6x standard Opus rates across the full context window (including requests over 200k input tokens). Pricing, rate limits, and access are the same as for Opus 4.6 fast mode; interested customers must join the waitlist. Fast mode does not support the Batch API and is not available for Claude Platform on AWS, Amazon Bedrock, Google Vertex AI, or Microsoft Azure Foundry — it is offered only via the direct Anthropic API (with the beta header) and supported native integrations (Claude Code, Cursor, Windsurf, v0, Warp). Fast mode first launched for Opus 4.6 in February 2026.

Anthropic blog "KPMG integrates Claude across its core business and workforce of more than 276,000 in strategic alliance" (2026-05-19) — Anthropic and Big Four firm KPMG signed a global strategic alliance to embed Claude across KPMG's core business and client-delivery platform, spanning more than 276,000 employees in 138 countries. Integration covers KPMG's flagship client platform Digital Gateway, tax and legal client tools, cybersecurity vulnerability detection and remediation, private-equity portfolio company modernization, the KPMG Blaze code modernization offering, and internal business-function modernization. Anthropic also launched KPMG Digital Gateway Powered by Claude with an initial focus on tax and private-equity clients; KPMG US will embed Claude into its PE product lines and co-develop Claude-powered products for portfolio companies. KPMG Global Chairman & CEO Bill Thomas said the alliance reflects "a shared commitment to responsible AI, prioritizing security, trust, and governance"; KPMG US Vice Chair (Tax) Rema Serafi said building tax-regulation adjustment tools that previously took weeks "now takes minutes". Financial terms were not disclosed; KPMG US had been adopting Claude for two years prior to the global alliance.

Anthropic blog "Widening the conversation on frontier AI" (2026-05-19) — Anthropic kicked off a cross-disciplinary dialogue initiative with scholars, clergy, philosophers, and ethicists drawn from 15+ religious and cross-cultural groups to discuss the moral formation of AI systems and how Claude's character and values are shaped; future cohorts will expand to legal scholars, psychologists, writers, and civic institutions. Anthropic also disclosed an experimental tool that lets Claude access ethical reminders mid-task, which in testing produced lower rates of misaligned behavior. The post argues that building safe, beneficial AI models requires the technical work of alignment, interpretability, safeguards, and evaluations — but that work isn't done (and AI isn't deployed) in a vacuum, so Anthropic wants broader wisdom-traditions dialogue to help shape AI's impact on work, institutions, and the distribution of power.

Elevated errors on Claude.ai and Claude Code login — status.claude.com posted a monitoring update at 04:46 UTC on 05-19 and marked it resolved at 04:51 UTC; the actual impact window was 04:34 to 04:42 UTC, about 8 minutes; users saw elevated request errors and latency within the Claude.ai applications and encountered errors when logging in (including authenticating Claude Code via Claude.ai); affected components were claude.ai and Claude Code; Anthropic did not disclose a root cause.

Claude Code 2.1.144 — adds `/resume` support for background sessions (sessions started via `claude --bg` or agent view now appear alongside interactive ones, marked `bg`); background subagent completion notifications show elapsed duration (e.g. "Agent completed · 3h 2m 5s"); `/plugin` browse and discover panes show when a plugin was last updated; `/model` now changes the model for the current session only, press `d` in the picker to set a default for new sessions; CLI copy renames "extra usage" to "usage credits" and `/extra-usage` to `/usage-credits` (old name still works). Improvements: fixed startup hanging up to 75s when `api.anthropic.com` is unreachable (captive portal, firewall, VPN) — side-channel API calls now time out after 15s; fixed garbled terminal output after a missed window-resize event (e.g. dragging a VS Code split-pane divider), now self-heals on the next frame instead of needing Ctrl+L; fixed progressive terminal display corruption in long sessions; reduced VS Code rendering glitches; fewer spurious tool errors during search — `head`/`tail` views satisfy the read-before-edit check and a no-match (exit 1) from `egrep`/`fgrep`/`git grep`/`git diff` is no longer a failure; improved recovery from pre-response stream stalls (retries streaming once instead of falling back to a slower non-streaming request); faster SDK/headless MCP startup (pre-wait overlaps startup, up to 2s faster with slow MCP servers). Bug fixes: macOS background sessions crashing with "exit 1 before init" under Full Disk Access-protected folders (regression in 2.1.143); unrecoverable conversation when reading an image file whose extension mismatches its contents (e.g. HTML saved as .png); `/branch` failing after entering a worktree or in some background sessions; Escape in the AskUserQuestion notes field aborting the turn; model selection not applying via IDE picker or `applyFlagSettings` after startup; resumed sessions picking up another session's `/model` choice; Bedrock/Vertex users unable to select "Opus (1M context)" (regression in v2.1.129); paginated MCP `tools/list` returning only the first page and silently dropping tools; MCP images with unsupported MIME types (e.g. SVG) breaking the conversation, now saved to disk; file descriptor exhaustion when a build runs in a skill directory; session title generated from plugin monitor output instead of the first prompt; Skill tool permission error in headless mode (regression in v2.1.141); `claude mcp list` silently reporting no servers when `.mcp.json` can't be parsed; scrolling in attached background sessions on Windows; crash when closing the terminal while attached to a background session; `! <cmd>` exec sessions not responding to Ctrl+C while attached.

Reuters reports that Anthropic has loosened the confidentiality rules around Project Glasswing, the program built on its Claude Mythos Preview model — partners were told roughly a week ago that they may now generally disclose their participation in Glasswing and, at their own discretion, share the vulnerability findings, best practices, tools, or code produced through the program with other companies' security teams, industry bodies, regulators, government agencies, open-source maintainers, the media, and even the general public, provided responsible-disclosure norms are observed. The original partner agreements had built-in confidentiality clauses because partners, worried about being targeted by attackers, asked for them; as the program matured, Anthropic eased the restrictions under pressure from lawmakers and others — critics argued that restricting access to vulnerability findings would put smaller organizations at a disadvantage. Glasswing partners include major tech firms such as Amazon, Microsoft, Nvidia, and Apple, and the Pentagon also deploys Mythos under a restricted program.

Elevated errors for Claude Opus 4.7 — status.claude.com began investigating at 22:10 UTC on 05-18 and resolved at 22:19 UTC; actual impact ran from 21:47 to 22:18 UTC, lasting about 31 minutes; affected components include claude.ai, Claude Console (platform.claude.com), Claude API (api.anthropic.com), Claude Code, and Claude Cowork; Anthropic did not disclose a root cause.

Anthropic acquires Stainless, a maker of SDK and MCP server tooling — the official blog announced the completed acquisition without disclosing deal terms. Founded in 2022, Stainless has generated every official Anthropic SDK since the API launched, turning an API spec into multi-language SDKs, CLIs, and MCP servers across TypeScript, Python, Go, Java, Kotlin, and more, and serves hundreds of enterprise customers. Head of Platform Engineering Katelyn Lesse said "agents are only as useful as what they can connect to," noting MCP makes agent connectivity possible and that combining the two teams strengthens Claude's data and tool integration; Stainless founder Alex Rattray is joining Anthropic.

Claude Haiku 4.5 elevated error rates — status.claude.com began investigating at 06:12 UTC on 05-18, deployed a fix and entered monitoring at 07:28 UTC, and resolved at 08:07 UTC, lasting roughly 1 hour 55 minutes; the impact spanned claude.ai, Claude Console (platform.claude.com), Claude API (api.anthropic.com), Claude Code, and Claude Cowork; Anthropic did not disclose a root cause.

Anthropic acquires Stainless — TechCrunch reports that, per The Information, the deal valued Stainless at more than $300 million (Anthropic did not confirm or disclose terms). Stainless is a New York developer-tools startup whose SDK generation is also used by rivals including OpenAI, Google, Cloudflare, Replicate, and Runway, so the acquisition effectively removes a key infrastructure supplier from competitors' hands. Anthropic confirmed to TechCrunch it will wind down all hosted Stainless products, including the SDK generator, but existing customers retain ownership of SDKs already generated and can freely modify and extend them.

Claude Developer Platform official release note (2026-05-12): Fast mode (research preview) now supports the Claude Opus 4.7 model. By setting `speed: "fast"`, `model: "claude-opus-4-7"`, and the `fast-mode-2026-02-01` beta header in an API request, developers get significantly faster output token generation (up to roughly 2.5x output tokens/sec), priced at 6x standard Opus rates across the full context window (including requests over 200k input tokens). Pricing, rate limits, and access are the same as for Opus 4.6 fast mode; interested customers must join the waitlist. Fast mode does not support the Batch API and is not available for Claude Platform on AWS, Amazon Bedrock, Google Vertex AI, or Microsoft Azure Foundry — it is offered only via the direct Anthropic API (with the beta header) and supported native integrations (Claude Code, Cursor, Windsurf, v0, Warp). Fast mode first launched for Opus 4.6 in February 2026.

Reuters, citing a Financial Times report, says AI startup Anthropic will brief the Financial Stability Board (FSB) on cyber vulnerabilities in the global financial system uncovered by its Mythos Preview model. The briefing was arranged at the request of FSB chair and Bank of England Governor Andrew Bailey, with the audience being leading finance ministries and central banks from FSB member states (G20 economies). Mythos — a cybersecurity model Anthropic announced last month but has not yet publicly released — is designed to detect long-standing vulnerabilities in web browsers, infrastructure, and software; Anthropic has previously claimed Mythos found "thousands of high-severity vulnerabilities, including some in every major operating system and web browser." Because of risk concerns, Mythos has been released to only around 40 organizations, mostly in the US, leaving companies and regulators elsewhere worried about uneven levels of protection. Regulators' core concern is that an AI built for cyber defense could instead be misused to mount more sophisticated attacks, particularly threatening a banking sector reliant on aging infrastructure. In the US, Treasury Secretary Bessent and Federal Reserve Chair Powell have met with major American banks to discuss emerging AI-linked cyber threats, and the Pentagon has confirmed deploying Mythos under a restricted program to patch vulnerabilities in US government systems. Reuters said it could not immediately verify the FT report; Anthropic and the FSB did not immediately respond to requests for comment.

Elevated error rates across multiple models — status.claude.com began investigating at 18:08 UTC on 05-16, implemented a fix and moved to monitoring at 18:18 UTC, and resolved at 18:24 UTC, lasting about 16 minutes; affected services span claude.ai, Claude Console (platform.claude.com), Claude API (api.anthropic.com), Claude Code, and Claude Cowork; Anthropic did not disclose a root cause.

Claude Code 2.1.143 — Adds plugin dependency enforcement: `claude plugin disable` now refuses when another enabled plugin depends on the target (with a copy-pasteable disable-chain hint), while `claude plugin enable` force-enables transitive dependencies; the `/plugin` marketplace browse pane now shows projected per-turn and per-invocation token cost estimates; new `worktree.bgIsolation: "none"` setting lets background sessions edit the working copy directly without EnterWorktree (for repos where worktrees are impractical); the PowerShell tool now passes `-ExecutionPolicy Bypass` (opt out with `CLAUDE_CODE_POWERSHELL_RESPECT_EXECUTION_POLICY=1`) and is enabled by default for Bedrock/Vertex/Foundry users on Windows (opt out with `CLAUDE_CODE_USE_POWERSHELL_TOOL=0`); `claude agents` now accepts `--add-dir`, `--settings`, `--mcp-config`, `--plugin-dir`, `--permission-mode`, `--model`, `--effort`, and `--dangerously-skip-permissions` flags. Improvements: background sessions preserve the model and effort level you set after waking from idle; Shift+Tab in attached agent sessions now includes auto mode in the cycle; `/bg` and `←`-detach preserve `--mcp-config`/`--settings`/`--add-dir`/`--plugin-dir`/`--strict-mcp-config`/`--fallback-model`/`--allow-dangerously-skip-permissions`; background sessions launched from `claude agents` honor `permissions.defaultMode` from settings.json; `claude --bg --dangerously-skip-permissions` persists across retire→wake; worktree cleanup no longer falls back to `rm -rf` when `git worktree remove` fails. Fixes: corrupt `.credentials.json` with a non-array `scopes` value hanging the CLI on startup or silently aborting OAuth token refresh; stop hooks that block repeatedly looping forever (now ends with a warning after 8 consecutive blocks, override via `CLAUDE_CODE_STOP_HOOK_BLOCK_CAP`); Esc/Ctrl+C not cancelling a pending `/loop` wakeup while Claude is idle; `/goal` evaluator firing while background shells or delegated subagents are still running; `NO_COLOR`/`FORCE_COLOR` in settings.json `env` stripping Claude Code's own UI colors; `/bg` without a prompt sending "continue" to the forked session; `--agent <name>` not finding plugin-contributed agents without the `plugin:` prefix; background agents false-positive worker-stall detection after host sleep or macOS App Nap; 5xx error messages pointing at status.claude.com instead of the configured gateway; background sessions silently capturing IDE file references into the warm spare's input; macOS background-job sessions getting "Operation not permitted" for files in `~/Documents`, `~/Desktop`, or `~/Downloads`; right-click paste in `claude agents` on Windows Terminal and WSL, agent view spawning repeated PowerShell processes on Windows when listing sessions, and stale-fragment rendering when scrolling attached background sessions on Windows Terminal; deleting a session from agent view not removing its transcript file; pressing ← in `claude agents` while a response streams leaving the agents list unresponsive; background daemon spawn not falling back to the running binary when the `~/.local/bin/claude` launcher is missing; and `claude agents --allow-dangerously-skip-permissions` defaulting to bypass mode instead of making it available in the permission cycle.

Yahoo Finance reports that Salesforce CEO Marc Benioff revealed on the All-In podcast that Salesforce projects spending roughly $300 million on Anthropic tokens in 2026, almost entirely on coding — Benioff called coding agents 'awesome' and said 'Anthropic is awesome,' crediting coding agents with unprecedented efficiency gains across service, support, distribution, and marketing, and with faster software deployment and sales of development work. Benioff also revealed Salesforce is building code-writing capabilities inside Slack so all employees can write code more easily. On cost control, Benioff argued a future 'intermediary layer' is needed to intelligently decide which token inputs go to frontier models like Claude and which can be routed to smaller, cheaper models, rather than sending every employee's tokens to a frontier model.

Claude blog post 'Deploying Claude across the legal industry' — an official Anthropic deployment guide for the legal sector, giving law firms and in-house legal teams product options, integration paths, and a phased adoption roadmap, building on the earlier Claude for Legal practice-area plugin expansion. The guide covers four product lines — Claude Chat, Claude Cowork, Claude for Microsoft 365, and Claude Platform — alongside 12 pre-built practice-area plugins spanning Commercial, Corporate, Employment, Privacy, Product, Regulatory, AI Governance, IP, and Litigation. Typical use cases include contract redlining, M&A due diligence, litigation preparation, privacy impact assessments, regulatory monitoring, and outside-counsel oversight, with a three-phase adoption plan to guide rollout. The post cites the FTI Consulting / Relativity General Counsel Report — 87% of general counsel now use generative AI (up sharply from 44% the prior year) — and uses Anthropic's own legal team as a case study, with a downloadable deployment guide attached.

Elevated error rates on requests to some models — status.claude.com began investigating at 00:18 UTC on 05-15, at 01:06 UTC identified the issue as affecting Claude Opus and Sonnet 4.6 with Opus 4.7 already back to normal, at 01:26 UTC both Sonnet 4.6 and Opus 4.7 returned to normal success rates with work continuing on Opus 4.6, resolved at 01:46 UTC, lasting about 1 hour 28 minutes; affected services: Claude API (api.anthropic.com) and Claude Code; Anthropic did not disclose root cause.

Reuters reporter Blake Brittain on the Anthropic $1.5B authors copyright settlement fairness hearing — U.S. District Judge Araceli Martinez-Olguin presided over the fairness hearing in San Francisco on 5/14 and declined to grant final approval, asking lawyers for more detail on multiple issues including "attorneys' fees and payments to lead plaintiffs"; no ruling date announced. The settlement covers over 480,000 works; authors and other copyright holders filed claims covering more than 92% of those works. A group of 25+ writers including Dave Eggers and Vendela Vida opted out of the settlement and filed a separate California complaint against Anthropic the same day. Authors sued Anthropic in 2024, alleging unauthorized use of pirated books to train Claude; predecessor judge William Alsup ruled training itself qualified as fair use but found Anthropic violated authors' rights by storing more than 7 million pirated books in a "central library". This is the largest known U.S. copyright settlement.

Claude Code 2.1.142 — Fast mode now defaults to Opus 4.7 (was Opus 4.6); set `CLAUDE_CODE_OPUS_4_6_FAST_MODE_OVERRIDE=1` to pin fast mode to Opus 4.6. Adds `claude agents` flags `--add-dir`, `--settings`, `--mcp-config`, `--plugin-dir`, `--permission-mode`, `--model`, `--effort`, `--dangerously-skip-permissions` for configuring dispatched background sessions; plugins with a root-level `SKILL.md` and no `skills/` subdirectory are now surfaced as a skill; the `/plugin` details pane and `claude plugin details` now show LSP servers a plugin provides; `/web-setup` warns before replacing an existing GitHub App connection; improved reactive compaction — the first summarize attempt now seeds from the original request's overflow size, avoiding a wasted near-full-context retry; configuring a prompt- or agent-type hook for `SessionStart`/`Setup`/`SubagentStart` shows a clear "use a command-type hook instead" error; removes the stale `/model claude-sonnet-4-20250514` suggestion from Usage Policy refusal messages. Fixes: `MCP_TOOL_TIMEOUT` not raising the per-request fetch timeout for remote HTTP/SSE MCP servers (was capped at 60 seconds); background sessions not recognizing pre-existing git worktrees, blocking Edit while EnterWorktree refused to create a duplicate; background sessions disappearing and daemon reconnect failing after macOS sleep/wake; daemon not exiting cleanly after binary upgrades (e.g., `brew upgrade`), causing dispatched agents to crash-loop; background agents crash-looping when Claude-in-Chrome extension is connected without a shared tab; clicking links in attached `claude agents` session; `claude agents` "v to open in editor" using daemon's default editor instead of `$EDITOR`/`$VISUAL`; `claude agents` deadlocking on Windows with network-drive working directories (Ctrl+C now works during startup); background-color bleed in Apple Terminal and other 256-color-only terminals; `claude --bg --dangerously-skip-permissions` not persisting across retire/wake; session titles being derived from URLs when the first message is a link; redundant `set_model` requests from remote clients injecting duplicate `/model` breadcrumbs; plugins using `skills: ["./"]` showing a false "path escapes plugin directory" error; plugin cache cleanup deleting the active plugin version directory; `/plugin` browse pane showing "0 installs" for newly published plugins; plugin advisories not naming every `plugin.json` key that shadows a default folder.

Elevated errors on Claude Opus 4.7 — status.claude.com began investigating at 20:33 UTC on 05-14, deployed a fix and entered monitoring at 21:07 UTC, resolved at 22:25 UTC, lasting about 1 hour 52 minutes; affected services: claude.ai, Claude API (api.anthropic.com), Claude Code, and Claude Cowork; Anthropic did not disclose root cause.

After Anthropic's 5/14 announcement splitting programmatic billing (`claude -p` and similar drawing from a separate Agent SDK credit pool from 6/15), the GitHub community responded with PTY-wrapper tools: `smithersai/claude-p` (v0.0.4, released 5/13, 279 stars, MIT) drives the interactive Claude Code TUI inside an in-process zmux PTY session — an ANSI scanner answers terminal capability queries while hooks intercept session start/completion events to inject prompts and extract results from the transcript; it calls itself a "drop-in replacement for `claude -p`" and states it exists "for educational purposes" to "demonstrate why client-side restrictions on how a product is used are fundamentally unenforceable." `Equality-Machine/claude-p` (v0.1.1, released 5/14, 37 stars, MIT) uses a similar pseudo-TTY approach, launching interactive Claude and reading `~/.claude/projects/**/<session-id>.jsonl` for output, with the tagline "use what you already paid for." Both route `claude -p`-equivalent automation through the interactive subscription quota instead of the new credit pool, but their developers explicitly call them developer-convenience tools that can break when the TUI changes — and circumventing the official billing split sits in a gray area of Anthropic's terms of service.

The Register reports Anthropic is splitting programmatic Claude usage (including `claude -p` headless mode, Agent SDK, Claude Code GitHub Actions, and third-party Agent SDK apps) out of subscription quotas; starting 2026-06-15, programmatic usage will draw from a separate monthly Agent SDK credit pool, with subscription fees converted to credits at API rates (more expensive than subscription pricing), unused credits forfeited monthly without rollover, and overages billed via an opt-in extra usage feature at pay-as-you-go API rates. Anthropic acknowledges prior enforcement gaps tracing to 2025 supply-demand imbalances, amplified by the February 2026 surge of OpenClaw, an open-source agent platform driving long token-intensive tasks. Interactive usage (claude.ai chat, Claude Code terminal interactive, Claude Cowork) remains on subscription.

Anthropic official support article "Use the Claude Agent SDK with your Claude plan" — documents the dedicated monthly Agent SDK credit launching 2026-06-15. Credit applies to: Claude Agent SDK in user-owned Python/TypeScript projects; `claude -p` in Claude Code (non-interactive mode); Claude Code GitHub Actions; third-party apps authenticated through Agent SDK. Monthly credits by plan: Pro $20, Max 5x $100, Max 20x $200, Team Standard $20/seat, Team Premium $100/seat, Enterprise (usage-based) $20, Enterprise (seat-based Premium) $200. Key rules: credits belong to individual accounts (no team pooling or sharing); unused credits do not roll over (reset at billing cycle start); one-time manual claim then automatic monthly refresh; Agent SDK usage draws from the monthly credit first, with overage going to standard API rates only if "extra usage" is enabled.

Bloomberg Law reporter Evan Ochsner on Anthropic's deeper push into legal work — analyzes Anthropic's May 12 release of 12 legal practice-area plug-ins and 20+ MCP connectors as a strategic shift from underlying model provider to direct competitor in legal-tech infrastructure; consultant Rudy DeFelice: "Rather than being the model that vertical tools rely on, Claude is positioning itself as the infrastructure through which work gets done"; analyst Chris Audet calls this rollout "bigger" than February's contract-redlining release that already shook legal-tech stocks; Zach Posner counters that specialists can still win on the final 20% of optimization through superior UX and deep domain expertise that horizontal players struggle to perfect; integrators named include Harvey, Relativity, Everlaw, Thomson Reuters; firms with live deployments include Freshfields, Quinn Emanuel Urquhart & Sullivan, Holland & Knight, Crosby Legal.