Daily AI Builder Brief - May 29, 2026

Executive Summary

Claude Opus 4.8 shipped May 28 and is now the default model in Claude Code 2.1.154. It scores 69.2% on SWE-bench Pro, up from 64.3% on Opus 4.7.
Dynamic Workflows landed in research preview: Claude writes a JavaScript script that orchestrates subagents in the background, capped at 1,000 per run.
Fast mode was repriced (roughly 2x the standard rate for about 2.5x the speed), and a lean system prompt is now the default.
Claude Code 2.1.152 made /code-review --fix apply findings directly, plus added disallowed-tools frontmatter, /reload-skills, and MessageDisplay hooks.
2.1.156 patched an Opus 4.8 bug where thinking blocks got modified and threw API errors.
Anthropic raised $65B at a $965B post-money valuation.
MCP's 2026-07-28 spec is locked as a release candidate: stateless core, Tasks, and MCP Apps. The final lands July 28.
Project Glasswing's Claude Mythos Preview found 10,000+ zero-days; a Cyber Verification Program opens the tooling to security teams.

Claude Code

Opus 4.8 is the default (2.1.154)

High-effort defaults, Dynamic Workflows, fast-mode improvements, and a lean system prompt all ship together. Re-baseline your agents.

Auto-fix reviews and new hooks (2.1.152)

/code-review --fix writes findings to your tree. New disallowed-tools frontmatter, /reload-skills, and MessageDisplay hooks expand control.

Thinking-block fix (2.1.156)

Resolves API errors on Opus 4.8 caused by modified thinking blocks. Update if you're on 4.8.

Anthropic

Claude Opus 4.8 (announcement)

Available across the API, Bedrock, Vertex AI, and Foundry at Opus 4.7 pricing. About 4x less likely to let flaws in its own code pass unremarked.

Dynamic Workflows (TechCrunch)

Script-driven orchestration of parallel subagents, for tasks one context can't hold.

$65B raise (US News)

Valuation hits $965B as Claude demand surges.

MCP Ecosystem

2026-07-28 release candidate (MCP blog)

Locked May 21, final July 28. The largest revision since launch: stateless at the protocol layer, so remote servers run behind a plain round-robin load balancer with no sticky sessions. Within the last 48 hours the ecosystem was quiet (SDK commits only), so validate now against the RC.

Broader AI

Glasswing finds 10,000+ zero-days (Anthropic, Help Net)

Claude Mythos Preview autonomously found 10,000+ high and critical zero-days across critical software. A Cyber Verification Program now hands the scanning, threat-modeling, and reporting tools to approved security teams.

What This Means For Builders

Re-test prompts and agents against Opus 4.8 defaults. High-effort mode and the lean system prompt change behavior, and the thinking-block fix is a hard prerequisite.
Dynamic Workflows plus cheaper fast mode shift the cost-and-throughput math on multi-agent jobs. Reach for them on fan-out work, not single edits.
If you ship MCP servers, port to the stateless RC before July 28 so you can drop sticky sessions.
Security shops: apply for Cyber Verification to get the Mythos vuln-research tooling.