Opus 4.8, dynamic workflows, and a $65B Series H

TLDR;

Opus 4.8 is the headline this week — new model, same price, stronger coding benchmarks, and a genuinely useful new feature for agentic loops (mid-conversation system messages). Claude Code also shipped dynamic workflows in research preview, letting you throw hundreds of parallel subagents at your biggest migrations — Salesforce reports a 231-day migration completed in 13 days, and Bun's 700k-LOC Rust rewrite shipped in 6. On the business side, Anthropic closed a $65B Series H at a $965B post-money valuation. Several hard deadlines are coming fast: the CLAUDE_CODE_OPUS_4_6_FAST_MODE_OVERRIDE env var dies June 1, and the claude-sonnet-4 / claude-opus-4 originals retire June 15.

Your AI shouldn't grade its own homework

Claude Code writes beautiful code. So does Codex. But here's the thing — they also think they write beautiful code. And when you ask an AI to review code it just wrote, you get the intellectual equivalent of a student grading their own exam. Shockingly, they always pass.

CodeRabbit CLI plugs into Claude Code and Codex as an external reviewer — different AI agent, different architecture, 40+ static analyzers, and zero emotional attachment to the code it's looking at. The agent writes, CodeRabbit reviews, and the agent fixes. Loop until clean.

You show up when there's actually something worth approving.

One command. Autonomous generate-review-iterate cycles. The AI still does the work. It just doesn't get to decide if the work is good anymore.

Free tier available. Try CodeRabbit's CLI →

API & Model Updates

Introducing Claude Opus 4.8

Released May 28 at the same price as Opus 4.7 ($5/$25 per million input/output tokens). SWE-bench Pro improves from 64.3 → 69.2; the model is notably more honest about uncertainty and catches its own bugs rather than declaring premature victory.

Mid-Conversation System Messages — New in Opus 4.8

You can now pass a role: "system" message mid-conversation in the messages array; Claude treats it as authoritative from that point forward, and it doesn't break the prompt cache. More cache hits = lower cost and latency on long agentic loops. Docs: prompt caching · mid-conversation system messages.

Opus 4.8 Fast Mode — Research Preview

Fast mode is now available for Opus 4.8 at 2× the standard rate for 2.5× the speed — roughly 3× cheaper than the previous Opus fast mode tier. Set /effort xhigh for your hardest tasks, fast mode for everything that needs throughput.

Opus 4.8 Prompt Caching — Lower Minimum Token Threshold

The minimum cacheable prompt length on Opus 4.8 drops to 1,024 tokens (down from Opus 4.7). Opus 4.8 also uses adaptive thinking, triggering reasoning only when a turn needs it — fewer wasted thinking tokens compared to running Opus 4.7 at the same effort level.

Opus 4.8: Drop-in Replacement for Opus 4.7 — No Breaking Changes

Same context window, same tool surface, same pricing. For teams already running Opus 4.7 in production, migration is a config-only model string swap. Run /claude-api migrate in Claude Code to update model strings and get prompting suggestions tuned for 4.8.

⚠️ claude-sonnet-4-20250514 and claude-opus-4-20250514 Retire June 15, 2026

Both original Claude 4 model snapshots are deprecated and retire on the Claude API on June 15. Migrate to Claude Sonnet 4.6 and Claude Opus 4.7 (or 4.8) respectively before that date.

⚠️ Claude Agent SDK Billing Split — Effective June 15, 2026

Programmatic Claude usage (Claude Agent SDK, claude -p, Claude Code GitHub Actions, third-party agents) is moving off the subscription rate-limit pool onto separate dollar-denominated credits at standard API list prices on June 15. Credit allocations: $20 (Pro), $100 (Max 5×), $200 (Max 20×) per month.

⚠️ SDK Package Rename: @anthropic-ai/claude-code → @anthropic-ai/claude-agent-sdk

Effective June 15: the TypeScript package @anthropic-ai/claude-code becomes @anthropic-ai/claude-agent-sdk; the Python package claude-code-sdk becomes claude-agent-sdk; ClaudeCodeOptions is renamed to ClaudeAgentOptions. Update your imports before the deadline.

⚠️ CLAUDE_CODE_OPUS_4_6_FAST_MODE_OVERRIDE Deprecated — Removed June 1

This env var is deprecated in Claude Code v2.1.154 and will be removed on June 1, 2026. Switch to /model claude-opus-4-6[1m] then /fast on to keep fast mode on Opus 4.6.

anthropic Python SDK v0.105.2

Released May 29, includes Opus 4.8 spec updates. v0.105.0 shipped May 28. Install: pip install anthropic==0.105.2.

Claude Code & CLI

Dynamic Workflows — Research Preview in Claude Code

Type "workflow" in a prompt to have Claude write an orchestration script on the fly and spin up tens to hundreds of coordinated subagents in parallel. Designed for migrations, large refactors, and performance optimization. Run /workflows to view your runs. Available on Max, Team, Enterprise, and via the API (Bedrock, Vertex AI, Foundry). Token-intensive — reserve for your biggest jobs. Bun's creator reported individual workflows running 10 hours continuously during their 700k-LOC Rust rewrite.

ultracode: Let Claude Decide When to Run a Workflow

Set /effort ultracode and Claude pairs xhigh reasoning with automatic workflow orchestration: it plans a workflow for each substantial task instead of waiting for you to type "workflow." A single request can become several workflows in a row (one to understand the code, one to make the change, one to verify it), so each request uses more tokens and takes longer. It is session-scoped and resets on a new session; drop back with /effort high for routine work. Available only on models that support xhigh effort. Shipped alongside dynamic workflows in Claude Code v2.1.154.

Claude Code v2.1.157 — Plugins Auto-load from .claude/skills

Released May 29. Plugins in .claude/skills directories are now automatically loaded — no marketplace required. New claude plugin init <name> scaffolds a plugin locally. Also: tool_decision telemetry events now include tool_parameters when OTEL_LOG_TOOL_DETAILS=1, and worktrees managed by Claude are left unlocked when the agent finishes so git worktree remove/prune can clean them up.

Claude Code v2.1.156 — Opus 4.8 Thinking Block Fix

Released May 29. Hotfix for an issue where thinking blocks were modified when using Opus 4.8, causing API errors. Update immediately if you're running 4.8 in Claude Code.

Claude Code v2.1.154 — Opus 4.8, Dynamic Workflows, Fast Mode

Released May 28. The flagship release this week: Opus 4.8 is now the default, dynamic workflows ship as a research preview, fast mode on Opus 4.8 is available at 2× cost/2.5× speed, and the lean system prompt is now default for all models except Haiku, Sonnet, and Opus 4.7 and earlier. Also: claude agents shell commands via ! <command>, /simplify now runs cleanup-only review rather than full bug-hunting, and effort slider labels renamed "Faster"/"Smarter".

Claude Code v2.1.153 — MCP Fixes, macOS Permissions, Git LFS Skip

Released May 28. Fixed stateful MCP servers without optional GET SSE stream reconnect-looping on tools/list (regression in v2.1.147). Fixed custom API gateways incorrectly receiving the user's Anthropic OAuth credential. Added skipLfs option to github/git plugin marketplace sources. macOS background agents now retain Privacy & Security grants across upgrades.

Claude Code v2.1.152 — /code-review --fix, Hook Improvements, Auto Mode No Longer Requires Opt-in

Released May 27. /code-review --fix now applies review findings directly to your working tree. Skills can set disallowed-tools in frontmatter. New /reload-skills command. SessionStart hooks can now return reloadSkills: true and set session titles. New MessageDisplay hook to transform or hide assistant message text. Auto mode no longer requires opt-in consent.

Security-Guidance Plugin for Claude Code — Shipped May 26

A new security plugin identifies and fixes vulnerabilities as you write code. Install via /plugins. Anthropic's internal rollout saw a 30–40% decrease in security-related PR comments. Add org-specific rules in a claude-security-guidance.md file in your repo or distribute via MDM.

Claude Code Reliability Update — May 27

Several quality-of-life fixes: fewer cryptic "tool result doesn't match tool use" errors with more readable messages; compaction now shows progress and no longer blocks on "prompt too long" bugs; more compaction speed improvements rolling out.

Agent SDK & Managed Agents

claude-agent-sdk-typescript v0.3.157

Released May 29. Parity with Claude Code v2.1.157 (auto-load plugins from .claude/skills, claude plugin init, EnterWorktree mid-session switching, tool_parameters in telemetry). Update: npm install @anthropic-ai/claude-agent-sdk@0.3.157.

claude-agent-sdk-typescript v0.3.154 — stdio MCP Fix

Released May 28. Fixed stdio MCP servers being incorrectly restarted on every reconcile pass due to config-equality false positives — a meaningful reliability fix for anyone using stdio MCP in SDK-driven agents. Update: npm install @anthropic-ai/claude-agent-sdk@0.3.154.

claude-agent-sdk-typescript v0.3.152 — Hook Events

Released May 27. SessionStart hooks can now return reloadSkills: true and set session title via hookSpecificOutput.sessionTitle. New MessageDisplay hook event for transforming or hiding assistant message text. Update: npm install @anthropic-ai/claude-agent-sdk@0.3.152.

claude-agent-sdk-python v0.2.87

Released May 23. Bundles Claude CLI v2.1.150; CI switched from static API key to Workload Identity Federation using short-lived tokens. Install: pip install claude-agent-sdk==0.2.87.

Developer Tools & Community

Salesforce: 231-Day Migration Completed in 13 Days with Claude Code

Published May 29. Salesforce published a detailed writeup on agentic Claude Code usage: one migration scoped at 231 days shipped in 13; one PR delivered 21 endpoints at 100% test coverage; total incidents dropped 5% even as PR volume increased. They built security guardrails directly into the agentic workflow. The core lesson: the teams seeing the biggest wins are redesigning their workflows entirely, not just speeding up existing ones.

Bun Rewritten in Rust in 6 Days Using Claude Code Dynamic Workflows

Published May 28. Follow-up to last week's "no behavior diffs after several days" report: Jarred Sumner confirmed the full 700k-LOC Bun → Rust rewrite shipped in just 6 days, with dynamic workflows and adversarial code review central to the result. He's seen individual workflows run for 10 continuous hours. The codebase is now under active fuzzing (using swc's test suite as corpus) to address security scrutiny.

How We Contain Claude Across Products

Published May 25. Anthropic engineering post on the isolation and containment architecture used when running Claude across different product surfaces — sandboxes, VMs, and egress controls to cap an agent's blast radius. Useful context for anyone designing multi-tenant Claude-powered applications.

Coding Agents in the Social Sciences

Published May 27. Research on using Claude coding agents for social science research tasks — relevant if you're building Claude-powered research automation tools or academic tooling.

Competitor Dev Tools

Cursor 3.6 — Auto-review Run Mode

Released May 29. New "Auto-review" run mode applies to Shell, MCP, and Fetch tool calls: allowlisted calls run immediately, sandboxable calls run in the sandbox, everything else goes to a classifier subagent that decides whether to proceed or ask. Configure in Settings > Cursor Settings > Agents > Run Mode.

GitHub Copilot: Claude Opus 4.8 GA — 15× Premium Multiplier Until June 1

Published May 28. Opus 4.8 is now generally available in GitHub Copilot for Pro+, Business, and Enterprise. Note: it launches at a 15× premium request multiplier until Usage Based Billing launches on June 1 — plan your usage accordingly if you're on a seat-limited plan.

Enterprise & Business

Anthropic Raises $65B Series H at $965B Valuation

Published May 28. Anthropic raised $65 billion at a $965 billion post-money valuation, led by Altimeter Capital, Dragoneer, Greenoaks, and Sequoia Capital, with run-rate revenue having crossed $47 billion earlier in May. The round funds safety and interpretability research, more compute for Claude, and scaling the products and partnerships customers rely on (Claude Code, Cowork). It includes $15 billion of previously committed hyperscaler investment ($5 billion from Amazon), plus new strategic memory and chip partners Micron, Samsung, and SK hynix.

Claude Marketplace — Five New Partners Added

Published May 27. Augment Code, Bolt, CodeRabbit, Hebbia, and Legora are now in the Claude Marketplace. Enterprise customers can apply existing Anthropic spend commitments toward these Claude-powered products. See claude.com/platform/marketplace.

Anthropic Opens Milan Office

Published May 27. New office to support Italian enterprise customers, researchers, and developers. Relevant if you're building Claude-powered products in the EU and want regional support contacts.

Research & Safety

Chris Olah's Remarks on Pope Leo XIV's Encyclical "Magnifica Humanitas"

Published May 25. Anthropic co-founder Chris Olah responds to the papal encyclical on AI and humanity. Background context on how Anthropic's interpretability work is being received in broader societal discourse.

Action Items

By June 1, 2026:

Remove CLAUDE_CODE_OPUS_4_6_FAST_MODE_OVERRIDE from your environment. Switch to /model claude-opus-4-6[1m] then /fast on for fast mode on Opus 4.6. (Deprecated in Claude Code v2.1.154, env var removed June 1.)
GitHub Copilot users: Claude Opus 4.8 launches at a 15× premium request multiplier until Usage Based Billing activates on June 1 — plan consumption accordingly.

By June 15, 2026:

Migrate off claude-sonnet-4-20250514 and claude-opus-4-20250514 — both retire on the Claude API on June 15. Target: Claude Sonnet 4.6 or Claude Opus 4.7/4.8.
Update SDK imports: @anthropic-ai/claude-code → @anthropic-ai/claude-agent-sdk (TypeScript); claude-code-sdk → claude-agent-sdk (Python); ClaudeCodeOptions → ClaudeAgentOptions. See the billing change and rename guide.
If your team uses claude -p, Claude Code GitHub Actions, or third-party Claude agents: the billing split moves programmatic usage to dollar-denominated credits on June 15 at standard API list prices.

Immediate (model migration):

Opus 4.8 is a drop-in for Opus 4.7 at the same price. Run /claude-api migrate in Claude Code to get updated model strings and Opus 4.8-tuned prompt suggestions.
If using Claude Code v2.1.154 or v2.1.155 with Opus 4.8: update to v2.1.156 to fix thinking block API errors.

Did you like this issue?

One tap is the fastest way to shape next week.

👍 Liked it · 👎 Didn't like it

And I read every reply. Hit reply and tell me what you want more of, or less of, and it goes straight into the next issue.