Claude Opus 4.8 Is Here — And New features

Anthropic’s newest flagship model is smarter, more honest, and now lets you decide how hard it thinks.

On May 28, 2026, Anthropic dropped a quiet but consequential upgrade: **Claude Opus 4.8**. No flashy rebrand, no new pricing surprise — just a model that’s noticeably sharper, more honest, and a lot more useful for the kind of long, messy, multi-step work that real people (and real engineers) actually do.

Here’s what’s new, why it matters, and what it means for builders, writers, analysts, and anyone whose job now includes “collaborating with an AI.”

TL;DR

• Claude Opus 4.8 launches with improvements across coding, reasoning, agentic, and knowledge-work benchmarks.
• Same price as Opus 4.7 — $5 / $25 per million input/output tokens.
• Effort control lets you choose how hard Claude thinks on each task.
• Dynamic workflows in Claude Code can now plan and run hundreds of parallel subagents for codebase-scale projects.
• Fast mode is now 3× cheaper than before.
• Honesty improvements: ~4× less likely to let flaws in its own code slip by unflagged.

What Actually Changed

1. It’s smarter where it counts

Opus 4.8 improves on Opus 4.7 across the board — coding, agentic skills, reasoning, and practical knowledge work. Anthropic published a full benchmark breakdown in the **Claude Opus 4.8 System Card**.

Image: Anthropic — Opus 4.8 benchmark comparison

But the more interesting story isn’t the numbers. It’s how the model behaves.

2. It’s more honest — and that’s a bigger deal than it sounds

One of the most underrated upgrades in modern AI is when a model stops pretending it finished something it didn’t. Anthropic claims Opus 4.8 is about four times less likely than its predecessor to let flaws in its own code pass without flagging them.

In practice, that means fewer moments of “the script ran, all good!” followed by you discovering it silently skipped half the work. Opus 4.8 is more likely to say, “Here’s what I did. Here’s what I’m unsure about. Here’s where you should look.”

That’s the difference between an assistant and a collaborator.

3. You’re now in charge of how hard it thinks

The new effort control sits right next to the model selector in claude.ai and Cowork. You pick:

• High (default) — balanced quality and speed.
• Extra (xhigh in Claude Code) — for difficult or long-running async tasks.
• Max — when you want it to spend whatever tokens it takes to get the right answer.
• Lower settings — when you want speed and want to conserve rate limits.

This is a small UI change that quietly shifts the relationship: instead of guessing what mode Claude is in, you set the intensity yourself.

4. Dynamic workflows: Claude Code goes big

This is the headline feature for developers. **Dynamic workflows let Claude Code plan a complex job, spawn hundreds of parallel subagents** in one session, verify outputs, and report back.

The example Anthropic gives: a codebase-scale migration across hundreds of thousands of lines of code, from kickoff to merge, with the existing test suite as the bar. That’s not “AI helps you refactor a file.” That’s AI taking on the kind of multi-week engineering ticket teams used to dread.

What Early Testers Are Saying

Anthropic shipped this with testimonials from teams across coding, legal, finance, and agent products. A few stood out:

“In Claude Code, it asks the right questions, catches its own mistakes, pushes back when a plan isn’t sound, and builds up confidence around complex, multi-service explorations before making big changes.” — Tom Pritchard, Staff Engineer

“On our Super-Agent benchmark, Claude Opus 4.8 is the only model to complete every case end-to-end, beating prior Opus models and GPT-5.5 at parity on cost.” — Kay Zhu, Co-Founder & CTO

“Claude Opus 4.8 delivers the highest score recorded on our Legal Agent Benchmark, and is the first model to break 10% overall on the all-pass standard.” — Niko Grupen, Head of Applied Research

“Claude Opus 4.8 is the strongest computer-use and browser-agent model we’ve tested, scoring 84% on Online-Mind2Web.” — Miguel Gonzalez, Tech Lead

“It finished faster and produced richer, more information dense outputs… The biggest differentiator was Opus 4.8’s tendency to proactively flag issues with the inputs and outputs of an analysis.” — Michael Ran, Sr. Investment Associate

The pattern across all the quotes: judgment, reliability, and self-awareness — not just raw capability.

Image: Anthropic

The Quiet Win: Alignment

Anthropic’s Alignment team noted that Opus 4.8 “reaches new highs on our measures of prosocial traits like supporting user autonomy and acting in the user’s best interest.”

Misaligned behaviour rates — things like deception or willingness to cooperate with misuse — are substantially lower than Opus 4.7, comparable to Anthropic’s best-aligned model so far, Claude Mythos Preview.

If you build products on top of Claude, this is the kind of thing that quietly translates into fewer support tickets, fewer surprises, and more trust from your users.

Also Launching Today

A few smaller updates that matter more than they look:

• Messages API now accepts system entries inside the messages array. Developers can update Claude’s instructions mid-task without breaking the prompt cache or routing the update through a user turn. Huge for agents that need to swap permissions, token budgets, or environment context on the fly.
• Fast mode is 3× cheaper than it was on previous models, and runs at 2.5× the standard speed.
• Effort control is available across all plans.

What’s Coming Next

Anthropic teased two things worth watching:

1. Cheaper models with Opus-level capability — they’re actively working on it.
2. A new class of model above Opus — codenamed Claude Mythos, currently used by a small set of organisations through Project Glasswing for cybersecurity work. General release is gated on stronger cyber safeguards, but Anthropic says it’s coming “in the coming weeks.”

My Take

The Opus 4.8 release isn’t about a single flashy capability. It’s about a model that finally feels like a steady collaborator — one that pushes back when your plan is shaky, flags its own uncertainty, and now lets you dial up the depth when you need it.

For developers, the dynamic workflows feature is the headline — codebase-scale migrations in one session is genuinely new territory. For everyone else, the honesty improvements and effort control are the things you’ll feel every single day.

If you found this useful, give it a clap, follow for more AI deep dives, and drop a comment with what you’d want me to try Opus 4.8 on next.