01What Claude Opus 4.8 Is and How It Works

Claude Opus 4.8 is Anthropic's newest large language model, focused on running AI agents at scale. Instead of one agent doing one task, Opus 4.8 can spin up hundreds of agents in a single session, all working on the same big job at the same time. It plans the whole job, splits the work among agents, and even checks its own work before saying it's done.

The key feature is called dynamic workflows. One lead agent coordinates helper agents, like a team of co-workers passing messages to each other. This lets you set a goal, walk away, and come back to find the job further along.

In this episode, Julian Goldie explains that this is the biggest upgrade Claude has ever shipped, and it changes how you use AI for real work.

02Main New Features of Claude Opus 4.8

Claude Opus 4.8 brings several new capabilities, but the standout is dynamic workflows. Here's what's new:

  • Hundreds of parallel agents: Run many agents at once on the same job, each taking a different piece.
  • Self-checking agents: The model reviews its own work before reporting completion, reducing mistakes.
  • Longer runtimes: Some jobs can stretch into days, meaning agents can work while you're away.
  • Effort controls: New settings in Claude chat and co-work let you dial effort from high to extra or max, or use fast mode.
  • Better memory: It keeps track of huge documents and long chat histories much better than before.
  • Less laziness: The new model is less likely to quit a task early.
  • Cheaper fast mode: Fast mode runs at 2.5 times the speed and is now three times cheaper.

There's also a new model coming called Miphos, which sits above Opus. Anthropic expects to bring Miphos-class models to all customers in the coming weeks, but it's not here yet.

03How to Set Up Claude Opus 4.8 for Agents

To get the most out of Claude Opus 4.8, you need to turn on dynamic workflows. Julian shows a simple setup in this episode. Here's how to start:

  1. Open Claude Code, the command-line tool for developers.
  2. Look for the Ultra Code mode and toggle it on. This turns on the highest effort setting plus dynamic workflows.
  3. Alternatively, type the phrase “create a dynamic workflow” in plain English in the chat.
  4. In the regular Claude chat, find the effort control and set it to extra or max for hard tasks, or lower it for simple ones.
  5. For a one-click setup, use the agent operating system inside the AI Profit Boardroom, which plugs in Opus 4.8 alongside other tools.
  6. Build your own workflow with prompts from the included template.

That's it. The setup is quick, and once it's on, you can start pointing swarms of agents at big jobs.

04How to Use Dynamic Workflows in Opus 4.8

Once dynamic workflows are on, using Opus 4.8 is all about giving it big goals and letting agents divide the work. In the episode, Julian demonstrates two setups: one lead agent that sends out helper agents, and a team of three or five agents that pass messages like co-workers, using send-message and wait-for-message tools.

For example, on a hard search task, a five-agent team beat a single agent and finished in about one fifth of the time. That's the core idea: the more agents you point at a tough job, the faster it gets. Julian also shows a simple build that took about one minute to complete, and notes that the output feels very human.

To use it, you can say “create a dynamic workflow” in your own words. The model will plan the job, split it into pieces, and assign agents. You can walk away and check back later, because some jobs can run for days.

Just set your effort level—high is default, but you can go to extra or max for harder tasks.

05Benchmark Results and Real-World Tests

Opus 4.8 shows strong numbers on many tests, though it's not a clean sweep. On SWE Bench Verified, real coding problems checked by human engineers, Opus 4.8 scores 88.6%, up from 4.7's 87.6%. On SWE Bench Pro, the harder version, it jumps to 69.2%, beating 4.7's 64.3%, GPT 5.5's 58.6%, and GMLive 3.1 Pro's 54.2%.

But on Terminal Bench, which tests coding in the command line, Opus 4.8 scores 74.6% (up from 66.1%), but GPT 5.5 still wins with 78.2%. On Terminal Work, GPT 5.5 also stays on top. So it's not a clean sweep.

On office work, Opus 4.8 beats GPT 5.5 by about 121 points on the GPT-PVAL test, winning roughly two out of three head-to-head matchups. On a legal test with over 1,200 tasks, it ranks top. On the Zapier test for business workflows, it scores 15.5%, up from 9.9%.

On the hardest long-context memory test, it scores 68.1% versus 4.7's 40.3%.

In a real-world test, developer Jared Sumner used dynamic workflows to rebuild the BUN project in a different language—750,000 lines in 11 days, with hundreds of agents and two reviewers per file. 99.8% of tests passed.

06Tips, Limits, and Mistakes to Avoid

One big tip from the episode: don't miss the effort setting. Opus 4.8 defaults to high effort, not the highest. You have two levels above that—extra and max—for tough jobs.

For simple tasks, turn it down to save your limits.

Another tip: use fast mode for speed, but know it runs at 2.5 times faster and is three times cheaper than before. But don't expect a huge change in everyday chat. Anthropic calls Opus 4.8 a modest but tangible improvement.

The big jump is on the agent side, not regular conversations.

Watch out for limitations: GPT 5.5 still wins on terminal coding tests, so Opus 4.8 isn't the best for every task. Also, the model is four times less likely to let flaws slide, but it's still not perfect. It might stop and say it's unsure, which is better than claiming false success.

Finally, Miphos is coming soon, and Anthropic says it expects to bring that class to all customers in coming weeks. So don't over-invest in 4.8 specifics if you plan to switch.

07Who Should Use Claude Opus 4.8

Claude Opus 4.8 is for anyone who wants to automate big, multi-step tasks—not just write a single prompt. Developers, project managers, and business owners can all benefit. If you have jobs that take a team of people weeks, you can now point a swarm of agents at them and let them run.

Julian emphasizes that the way we work with AI is shifting from one prompt at a time to setting a goal and letting a team of agents handle it. People who learn to set up and run agent swarms will get a lot more done. That's why he built the agent operating system in the AI Profit Boardroom, which includes a zip file, video tutorial, 30-day roadmap, and daily updates.

There are also weekly coaching calls and a community of 3,200 people.

If you don't code, you can still use dynamic workflows by typing plain English like “create a dynamic workflow.” The model handles the planning and execution. So even non-developers can leverage hundreds of agents for office work, legal tasks, or business automations. Just test it on your own, play with effort levels, and keep an eye out for Miphos.