Anthropic shipped its biggest release of H1 2026 today. Claude Opus 4.8 is generally available with dynamic workflows, adaptive thinking, a 1M token context window, and what early testers call a meaningfully more honest model. That lands in the same week Anthropic reported its first-ever quarterly operating profit ($559M on $4.8B Q1 revenue), launched the Claude Compliance API with 28 enterprise integrations, and posted the first update from Project Glasswing (10,000+ critical vulnerabilities found). Andrej Karpathy joined for pre-training research. This is the Claude Weekly for May 28, 2026. Read last week’s edition.

Claude Opus 4.8 β€” What Actually Changed

The model is available today on claude.ai, the Claude API, Amazon Bedrock, and Vertex AI. GitHub Copilot added it at GA as well. Same price as Opus 4.7.

The headline capability: Opus 4.8 is more honest. Early testers consistently report it’s more likely to flag uncertainty about its own work and less likely to confidently claim progress when the evidence is thin. For any workflow where humans rely on Claude’s self-assessment β€” code review, agentic task completion, research summaries β€” this is the upgrade that matters most. (Anthropic)

Technical changes on claude.ai:

  • Effort control slider β€” New UI control to adjust task effort. Higher settings trigger deeper reasoning; lower settings return faster responses using less rate-limit quota.
  • 1M token context window by default (128k max output tokens).
  • Adaptive thinking β€” Extended reasoning fires only when the turn warrants it. Simple lookups don’t burn thinking tokens; complex multi-step reasoning does.

Benchmark results: Opus 4.8 is the only model to complete every case end-to-end on Anthropic’s Super-Agent benchmark at GPT-5.5 parity cost. It scores 84% on Online-Mind2Web (browser-agent tasks), ahead of both Opus 4.7 and GPT-5.5. On CursorBench it exceeds Opus 4.7 at every effort level. In legal AI, it’s the first model to break 10% on the all-pass standard of the Legal Agent Benchmark β€” a number that directly translates into how much substantive attorney work can be delegated with confidence. (Anthropic)

Claude Code: Dynamic Workflows + Fast Mode Overhaul

The Claude Code release shipping today alongside Opus 4.8 is one of the most feature-dense updates since the tool’s launch.

Dynamic Workflows (Research Preview): Ask Claude to create a workflow and it orchestrates the task across tens to hundreds of agents running in the background. /workflows lists active runs. For complex multi-repo or multi-service engineering tasks that previously required a human breaking down work manually, this is the unlock. (TechCrunch)

Fast mode repriced: Fast mode on Opus 4.8 runs at 2.5Γ— speed for 2Γ— the standard rate β€” roughly 3Γ— cheaper than fast mode was on Opus 4.6. Max plan users default to fast mode on Opus 4.8. Claude Code now uses the lean system prompt by default for all models except Haiku, Sonnet, and Opus 4.7 and earlier. (GitHub Releases)

Agents view: Type ! in the agents view to run a shell command as a background session you can attach and detach from (also available as claude --bg --exec ''). The ←← shortcut to open agents view now works on Bedrock, Vertex, Foundry, and with telemetry disabled.

Chrome integration: Pick which connected browser to use via /chrome β†’ "Select browser…" β€” useful when multiple browsers are connected.

Plugin system maturation:

  • Plugins can declare defaultEnabled: false in plugin.json, enabling opt-in installation.
  • The /plugin Discover tab pins contextually relevant plugins with a “suggested for this directory” annotation.
  • Official plugin marketplace at github.com/anthropics/claude-plugins-official β€” install via /plugin install {name}@claude-plugins-official.

MCP in Claude Code: Stdio MCP server subprocesses now receive CLAUDE_CODE_SESSION_ID and CLAUDECODE=1. claude mcp list/get shows unapproved .mcp.json servers as ⏸ Pending approval (was silently auto-approving on piped output).

Deprecation watch: CLAUDE_CODE_OPUS_4_6_FAST_MODE_OVERRIDE is deprecated, removed June 1. Migrate to /model claude-opus-4-6[1m] + /fast on. (GitHub Releases)

Cloud-based Routines: Claude Code now supports three scheduling modes β€” cloud-based Routines (run when your computer is off), Desktop scheduled tasks (app must be open), and cron-style. The Routines mode is drawing attention as a replacement for no-code automation tools. (Routines Guide)

Claude API: Mid-Conversation System Messages + stop_details

Mid-conversation system messages (Opus 4.8 only): The Messages API now accepts role: "system" entries inside the messages array after a user turn, subject to placement rules. This allows developers to update Claude’s instructions mid-task without breaking the prompt cache or losing conversation context. No beta header required. (API Docs)

This is a meaningful ergonomic improvement for long-running agentic sessions β€” previously, updating system instructions mid-task required either a full session restart or careful workarounds to preserve cache hits.

stop_details on refusal responses (public): The stop_details field is now documented. It returns category ("cyber", "bio", or null) and a human-readable explanation. Applications can now route different refusal classes to appropriate handling (logging, escalation, user messaging) without string-matching on response text. No beta header required. (API Docs)

Prompt caching improvements: Minimum cacheable prompt length drops to 1,024 tokens on Opus 4.8 (lower than 4.7). With adaptive thinking enabled, Opus 4.8 also wastes fewer thinking tokens than 4.7 at the same effort level, improving cost efficiency for reasoning-heavy pipelines. (API Docs)

Cache diagnostics (public beta, since May 13): Pass diagnostics.previous_message_id and the API returns a cache_miss_reason explaining where the prefix diverged. Useful for debugging unexpected cache misses in agentic loops. Requires the cache-diagnosis-2026-04-07 beta header. (API Docs)

Claude Agent SDK & Managed Agents: Enterprise-Grade Expansion

Self-hosted sandboxes (public beta, since May 19): Claude Managed Agents can now execute tools in a sandbox you control β€” Cloudflare, Daytona, Modal, and Vercel as launch providers. This moves the primary enterprise blocker for production agentic deployments (data isolation) from “ask Anthropic” to “configure your own infrastructure.” (API Docs)

MCP tunnels (Research Preview, since May 19): Outbound-only encrypted connections that let agents reach MCP servers in private networks without inbound firewall changes. The first structural change to MCP’s production deployment model since OAuth 2.1 formalization. (API Docs)

Multiagent sessions & Outcomes (public beta, since May 6): Available under managed-agents-2026-04-01 beta header. Webhooks for session and vault lifecycle events. Large tool outputs >100K tokens now spill to a sandbox file automatically, with the model receiving a truncated preview and the file path. (API Docs)

Claude Platform on AWS (since May 11): Full Messages API, Files API, Message Batches API, Claude Managed Agents, Agent Skills, code execution, and tool use through native AWS endpoints with AWS billing and IAM auth. (API Docs)

MCP Ecosystem: 15,930+ Servers, AWS at 60+, Salesforce Integration

The MCP ecosystem has reached significant scale. Digital Applied’s May 2026 tracker catalogs 56 production-ready vendor-backed servers across 10 categories β€” a fraction of the full ecosystem that now spans 15,930+ indexed servers on PulseMCP, ~7,300 on Smithery, and ~2,000 on the official MCP Registry. (Digital Applied)

AWS Agent Toolkit: AWS shipped general availability with 60+ official MCP servers covering documentation, infrastructure, Lambda, containers, AI/ML, and analytics β€” the largest first-party catalog from any cloud provider.

Salesforce Hosted MCP Servers: Salesforce published integration documentation for connecting Claude Code directly to the Salesforce Platform, with both standard and custom hosted servers supported.

Observability expansion: IBM Instana, Honeycomb, and New Relic all shipped production MCP servers in May, growing the observability category from 6 to 9 vendor-maintained servers in 60 days.

For enterprise teams standing up MCP infrastructure, the deployment model has fundamentally changed: hosted endpoints, OAuth 2.1, and outbound-only gateways (via MCP tunnels) are now the standard shape, not the advanced configuration.

Anthropic’s First Profitable Quarter

This is the story of the week β€” even alongside the Opus 4.8 launch. Anthropic reported $4.8 billion in Q1 2026 revenue and its first-ever quarterly operating profit of $559 million. The company projects $10.9 billion in Q2 2026, implying an annualized run rate approaching $44 billion β€” which would rank Anthropic as the top private AI firm by revenue.

The profitability milestone is significant for enterprise buyers in a way that press coverage often misses: a profitable Anthropic is a structurally stable Anthropic. The existential “will they survive long enough for us to build on their APIs” concern that quietly appeared in enterprise procurement conversations six months ago is effectively off the table. (Crypto Briefing, The Information)

Bloomberg reported separately that Anthropic is weighing an IPO as early as October 2026 β€” likely its last major private fundraising event. Combine that with the SpaceX Colossus 1 compute partnership (structured as an initial 180-day lease for 300+ megawatts, confirmed by Musk on May 28) and you have a company consolidating its infrastructure position ahead of a potential public market debut.

Claude Compliance API: 28 Integrations at Launch

The Claude Compliance API launched May 21, giving enterprise security teams programmatic access to audit data on Claude usage across Anthropic’s product suite. At launch, Anthropic announced 28 third-party integration partners β€” a number that, in prior enterprise platform launches, has typically been a 12-to-18-month buildout. Doing it at GA is a statement.

Key integrations worth noting for enterprise teams:

  • Microsoft Purview β€” detect and investigate Claude usage alongside other cloud apps (Microsoft)
  • Okta β€” monitor and remediate identity risks within Claude (Okta)
  • Cloudflare CASB β€” Zero Trust extension to monitor Claude Enterprise activity from the Cloudflare dashboard (Cloudflare)
  • SailPoint β€” identity security posture management for Claude (SailPoint)
  • Proofpoint β€” data protection and insider risk detection (Proofpoint)
  • Relativity β€” collect Claude Enterprise data in RelativityOne for eDiscovery (PR Newswire)
  • Sumo Logic β€” track Claude alongside other enterprise logs for compliance investigations (Security Brief AU)

The broader point: enterprises have spent years building security and audit infrastructure around their SaaS stack. The Claude Compliance API slots Claude into that existing governance layer rather than requiring a parallel one. For organizations where the security team is the bottleneck on AI approval, this removes a significant blocker. (HelpNet Security)

Project Glasswing: 10,000+ Vulnerabilities Found

Anthropic published the first update on Project Glasswing β€” a multi-org initiative (AWS, Apple, Cisco, CrowdStrike, Google, JPMorgan, Microsoft, NVIDIA, and others) to proactively find and disclose vulnerabilities in critical open-source software using AI.

Using “Claude Mythos Preview” β€” a specialized security-focused model not available for general release β€” the initiative has identified more than 10,000 high- or critical-severity vulnerabilities in the most systemically important open-source projects in the world. As of May 22, 1,596 have been publicly disclosed and 97 patched. (Anthropic Research)

The “Claude for Open Source” program launched alongside it, supporting open-source maintainers and contributors. Claude Security entered public beta for Enterprise customers.

This matters beyond the vulnerability count. Glasswing is Anthropic demonstrating a use case β€” AI-assisted offensive security research at scale β€” that no human team could execute at that volume. For security and infrastructure teams evaluating where AI delivers asymmetric returns, this is a concrete proof point.

Andrej Karpathy Joins Anthropic

Andrej Karpathy β€” AI researcher, OpenAI co-founder, and creator of the widely-used Neural Networks: Zero to Hero course β€” joined Anthropic on May 19 to work on pre-training research. (MindStudio)

Pre-training hires at this level signal a specific organizational intent: Anthropic is competing on foundational model architecture, not just fine-tuning and product velocity. Combined with the Colossus compute access and first-profitable-quarter financials, the company has the resources and the talent trajectory to accelerate at the model level. Watch for architectural announcements in H2 2026.

Anthropic launched a Claude for the Legal Industry initiative this week, driven by in-house counsel Mark Pike. The program addresses governance, compliance, and professional responsibility requirements specific to legal workflows β€” the areas that have kept large law firms in evaluation mode rather than deployment mode for the past year. (Law360)

Combined with Opus 4.8’s Legal Agent Benchmark result (first model to break 10% on the all-pass standard), the legal vertical is receiving a coordinated push from both the model and the go-to-market side.

Benchmarks: Opus 4.7 Leads SWE-Bench at 82%

Third-party benchmark tracker Vals.AI (updated May 25) places Claude Opus 4.7 at 82.00% on SWE-bench Verified β€” ahead of Gemini 3.1 Pro Preview (78.80%) and GPT-5.5. A separate comparison from SwiftWand puts Opus 4.7 at 80.8% with a continued lead on ARC-AGI-2 at roughly 32%. (Vals.AI)

Note: these are Opus 4.7 numbers, the pre-4.8 baseline. Expect new SWE-bench submissions for Opus 4.8 in the coming days.

Community & Ecosystem

VsAgentic β€” new community extension for Visual Studio 2026, hooking into the Claude Code CLI for in-IDE agentic coding assistance.

awesome-claude-plugins β€” now tracking and ranking the top 100 Claude Code plugin repositories across an index of 18,697+ total repositories.

OpenTelemetry observability for Claude Code β€” General Analysis published a practical guide to Claude Code observability with OpenTelemetry, including metrics, events, traces, tool decisions, hooks, and MCP activity. (General Analysis)

Anthropic Engineering Blog: “How we contain Claude across products” β€” a detailed post on containment strategies across products, noting that Claude Mythos Preview was deemed too high a blast radius for broader release. Between mid-2025 and January 2026, Anthropic received and addressed three responsible disclosure vulnerability reports for Claude Code. (Anthropic Engineering)

Competitive Snapshot

Google I/O 2026 was the week’s primary competitive storyline. Gemini 3.5 Flash launched at approximately 2Γ— cheaper on input and 40% cheaper on output compared to Claude Sonnet β€” a direct cost-pressure play for high-volume use cases. Google also unveiled Gemini Spark, an autonomous agent at $100/month, and showed prototype Android XR AI glasses. (TechCrunch, AI Builder Club)

OpenAI’s GPT-5.5 Pro sits at $30/$180 per million tokens β€” considerably more expensive than Opus β€” keeping Anthropic’s frontier pricing competitive. Meta launched paid subscriptions for Instagram, Facebook, and WhatsApp with AI tiers coming, entering the consumer AI subscription market directly. (dev.to pricing roundup)

What to Watch

  • Opus 4.8 Dynamic Workflows in production β€” The research preview label means API-available but not production-hardened. Watch for GA timing and how it changes cost profiles for complex agentic pipelines.
  • Agent SDK billing split effective June 15 β€” Programmatic usage (Agent SDK, claude -p, third-party harnesses) moves to separate per-plan credit pools. Teams running agentic workflows should audit usage patterns before the cutover.
  • CLAUDE_CODE_OPUS_4_6_FAST_MODE_OVERRIDE removal June 1 β€” If you’re using this env var, migrate now.
  • Anthropic IPO timeline β€” An October 2026 debut would likely be the last window to access Anthropic’s equity at pre-public pricing. Watch for formal filing signals.
  • Claude Compliance API ecosystem expansion β€” 28 integrations at GA is aggressive. Expect SIEM additions (Splunk, Microsoft Sentinel) in the next cycle.
  • Karpathy’s impact on pre-training β€” The combination of Colossus compute, first profitable quarter, and a foundational model hire suggests meaningful architectural news is coming in H2.

That is this week’s Claude Weekly. The Opus 4.8 launch alone would dominate any normal week. The combination of a profitable Anthropic, the largest enterprise compliance expansion in the company’s history, MCP tunnels enabling private-network deployments, and a foundational model hire of Karpathy’s caliber makes this one of the most consequential seven-day stretches in the company’s history.

If your team is evaluating Claude Enterprise, mapping compliance integrations to your security stack, planning for the June 15 billing changes, or building on Claude Managed Agents, reach out to Big Hat Group. We specialize in AI automation and agent deployment for enterprise environments β€” we’ll help you move from evaluation to production.

Check back next week.