Claude Opus 4.8 arrives but the quieter updates matter more

What Anthropic launched

On 28 May 2026 Anthropic released Claude Opus 4.8. It improves on Opus 4.7 across coding, reasoning and agentic work. The price stays the same.

Most coverage will stop at the model. That is the headline. Two changes launched alongside it will shape how your team works far more.

The first is dynamic workflows in Claude Code. The second is a new effort control in the Claude app and Cowork. Both are quieter than a version number. Both matter more.

Dynamic workflows hand Claude much more agency

Claude Code is arguably Claude's most capable form. Dynamic workflows make it stronger again.

Claude now writes its own orchestration scripts. It breaks a large job into pieces. It runs tens to hundreds of subagents in parallel. It checks the work before anything reaches you.

The checking is the clever part. Some agents attack the problem from independent angles. Others try to break what the first ones found. The run repeats until the answers agree.

Anthropic points to real uses. Codebase-wide bug hunts. Migrations across hundreds of files. Plans stress-tested from every angle before you commit.

One example stands out. A developer used dynamic workflows to rewrite the Bun runtime from one programming language to another. The result was roughly 750,000 lines of code with 99.8% of the test suite passing. It took 11 days.

This is a real step up in what one session can do. It is also a real step up in autonomy.

More autonomy raises the stakes

When a tool takes on more agency, the discipline around it matters more, not less.

Anthropic is candid about this. Dynamic workflows use far more tokens than a normal session. The company recommends starting on a small, scoped task. Claude Code shows you what it plans to run and asks you to confirm first.

Read that as a signal. Power without a plan is a liability. Our view at gecco has not changed. Strong results come from three things. A clear plan. Strong guardrails. A human who stays in the loop.

There is one practical point we will repeat for as long as it is true. If you use the API, set a spend limit on your account. Autonomous tools that burn tokens can burn money just as fast. A cap is the cheapest insurance you will buy.

Opus 4.8 helps here too. Anthropic says it is around four times less likely than the old version to let code flaws slip through. It also flags uncertainty more readily. Better self-checking is a guardrail in its own right. It does not replace yours.

Effort control is useful and one knob too many

The second change is a new effort control. It sits next to the model selector in the Claude app and Cowork. You choose how hard Claude works on a task.

Higher effort means deeper thinking and better answers. Lower effort means faster replies and slower use of your limits. Opus 4.8 defaults to high. You can push it to extra or max for difficult work. The control is available on every plan.

The feature is sensible. The problem is the pile of choices it now sits on.

We onboarded a client this week. Teaching models is already hard. Opus for this, Sonnet for that, Haiku for the other. Then you add adaptive thinking. Now you add effort levels on top.

For a business just starting out, that is a lot of dials. Every dial is a chance to get it wrong, or to freeze and do nothing.

This is where the auto mode in ChatGPT has an edge. It picks a sensible setting for you. Most SME users do not want to tune a model. They want a good answer. Sensible defaults beat more options for most people, most of the time.

The gap that actually matters for SMEs

Here is the part the headlines miss. Most of this new power lives in Claude Code. Claude Code is a developer tool.

Many teams we meet have not made the easy moves yet. Many are still on an early Copilot rollout that never quite took hold. They have not tried Claude in the browser. They have not opened Cowork. Claude Code is not even on the horizon.

So the frontier races ahead while everyday adoption stands still. That gap is the real story. It is not a technology problem. It is a people and process problem.

This is the split we talk about constantly. AI adoption is 80% people and culture, 20% technology. A more capable model does nothing for a team stuck on tools it has not learned to use.

What to do this week

• If you use the API, set a hard spend limit today. Do it before you test anything new.
• Pick one repetitive task and run it in Claude or ChatGPT in the browser. Build the habit before you chase the frontier.
• Agree a simple default for your team. One model, one effort setting, written down. Remove the choice for everyone who does not need it.
• Write a short plan before any large AI task. State the goal, the limits and who checks the output.

Where this leaves you

Opus 4.8 is a solid upgrade. Dynamic workflows are a genuine leap for complex work. Effort control is a useful addition.

None of it helps if your team has not cleared the first hurdle. The advantage now goes to businesses that adopt well, not the ones that own the newest model.

That is the work gecco does. We help UK SMEs build the plans, guardrails and habits that turn AI tools into measurable results. We start small and scale when you are ready. Take the AI readiness assessment to see where your team stands.

Sources

• Anthropic, Introducing Claude Opus 4.8. Published 28 May 2026.
• Claude, Introducing dynamic workflows in Claude Code. Published 28 May 2026.