What's New in Claude Opus 4.8 and Claude Code
Anthropic shipped Claude Opus 4.8 in late May 2026 — just over a month after Opus 4.7 — and paired it with a meaningful round of Claude Code updates. The theme is consistent: more capable agents, more control over how hard (and how expensively) they think, and tooling built for production-scale work rather than demos. Here's a practical rundown of what changed and why it matters.
Claude Opus 4.8: the model
More capability at the same price
Anthropic positions Opus 4.8 as its most capable generally available model to date, with stronger coding and knowledge-work performance — at the same price as the previous version. On Anthropic's reported agentic benchmarks it's the standout on long, end-to-end tasks. As always, treat vendor benchmarks as a starting point and validate against your own workloads, but the direction is clear: the frontier keeps moving on multi-step, tool-using work.
Adaptive thinking and effort controls
The most useful day-to-day change is effort control. The effort parameter now defaults to high across surfaces — the Claude API and Claude Code alike — and Opus 4.8 uses adaptive thinking: it reasons only when the turn warrants it. Simple lookups and short agentic steps get a direct answer; complex, multi-step problems trigger extended reasoning first. In practice that means less latency and cost on the easy stuff without giving up depth on the hard stuff — and a knob you can turn down deliberately when you want speed over deliberation.
A faster, cheaper "fast mode"
For interactive work, fast mode on Opus 4.8 runs at roughly 2.5× the speed and is about
three times cheaper than on previous models. In Claude Code you toggle it with
/fast — and importantly, it's still Opus, just optimized for faster output, not a downgrade to
a smaller model. That changes the economics of keeping a top-tier model in the loop for routine edits and
quick iterations.
A 1M-token context window
Opus 4.8 offers a 1,000,000-token context window, enough to hold large codebases, long document sets, or extended agent sessions in a single context. Combined with the agentic tooling below, it's what makes whole-repository work practical rather than a juggling act of partial views.
Better-aligned, not just more capable
Anthropic also reports alignment gains: higher scores on prosocial traits like supporting user autonomy and acting in the user's best interest, and lower rates of misaligned behavior such as deception than Opus 4.7. That matters more, not less, as we hand agents real privileges — a theme we've written about in Security in the Age of AI.
Claude Code: the tooling
Dynamic workflows for large-scale work
The headline Claude Code feature is dynamic workflows — orchestration built for large-scale problems, capable of migrating codebases of hundreds of thousands of lines. Rather than one long, fragile conversation, a workflow decomposes the job into stages that can fan out across many subagents and recombine — the structure you want for migrations, audits, and sweeping refactors where a single context can't (and shouldn't) hold everything.
Subagents and agent teams
Subagents are specialized instances that run in their own context window; their verbose output stays isolated and only a summary returns to the main conversation. Agent teams formalize this: a main agent owns planning and integration while specialist subagents — each with its own prompt and tool permissions — handle bounded tasks. The per-subagent permission scoping is the part worth dwelling on: it's least privilege applied to your AI, exactly the discipline that keeps an over-eager agent from doing damage.
Skills
A skill is a SKILL.md file under .claude/skills/<name>/,
invoked explicitly with /skill-name or automatically when Claude judges it relevant. Skills
run in the current context (no new process, no spawning), which makes them a lightweight way to package
repeatable, domain-specific procedures. A /reload-skills command re-scans skill directories
without restarting the session.
Hooks — including real security checkpoints
Hooks fire at a couple dozen lifecycle points and let you run your own logic around the agent. Two are
especially relevant to governance: UserPromptSubmit can block or modify a prompt before
Claude sees it, and PreToolUse runs before any tool executes — a natural place to gate or
deny dangerous actions. Skills and slash commands can also declare disallowed-tools to remove
capabilities while active. These are the building blocks for putting guardrails around what an agent is
allowed to do, in your own environment.
Quality-of-life: /usage and /code-review --fix
Recent updates also include a /usage breakdown that shows what's actually driving your limits
— skills, subagents, plugins, and per-MCP-server cost — and /code-review --fix, which applies
review findings (reuse, simplification, efficiency) directly to your working tree after the review.
The takeaway
Opus 4.8 and the latest Claude Code make agentic engineering faster, cheaper, and capable of genuinely large jobs — whole-repo migrations, multi-agent reviews, repeatable skills. The same release also leans harder into control: adaptive thinking and effort dials, per-subagent tool permissions, and hooks that can gate actions. That pairing is the right shape. As we argued in Why certifications still matter, the value isn't in generating the work — it's in the judgment to direct, verify, and contain it. The more powerful the agent, the more that judgment is worth.