Just 43 days after Opus 4.7 shipped, Anthropic has released Claude Opus 4.8. This is not a simple benchmark bump — it takes targeted aim at the two pains developers complain about most: AI laziness and overconfidence. The Dynamic Workflows feature launching at the same time lets Claude Code spin up hundreds of subagents at once to finish tasks in parallel.

Three Key Improvements in Opus 4.8

1. More honest, no more laziness

Anyone who has written code with Claude knows the pattern: it rattles off a big chunk of code, confidently declares everything works. You run it, and something else breaks. You go back and ask; it says it found the issue, definitely fine now. You run it again — another error.

Opus 4.8 is specifically optimized for this. Official numbers:

• The code-defect miss rate is down to 1/4 of Opus 4.7's
• The rate of overconfident behaviors like "hardcoding answers" is down to 1/10 of Opus 4.7's
• On laziness-detection metrics, Opus 4.8 is the only model achieving a 0% bad-behavior rate

In hands-on testing, Opus 4.8 reviews code thoroughly and in great detail, hunting for every spot that could use optimization, instead of wrapping up hastily like the previous generation.

2. More precise, but also more "obedient"

Opus 4.8 has become more precise — it goes exactly where you point and follows instructions more closely. For professional developers that is a good thing -- both error rates and hallucination rates are falling.

But it brings a change: its initiative has diminished. Ask it to do A, and now it does only A — it will no longer take it upon itself to also handle B. If your habits depend on the AI's "guess what you meant" ability, you will need to adjust how you interact and state requirements more explicitly.

3. Effort control open to all users

Every plan (including free users) can now adjust the model's effort level in Chat mode, across four settings from Low to Max. Combined with adaptive thinking, you can choose flexibly based on task complexity.

Dynamic Workflows: Hundreds of Subagents in Parallel

Dynamic Workflows, launched the same day as Opus 4.8, is a major Claude Code upgrade, currently available as a research preview in the CLI, the desktop app, and the VS Code extension.

How it works

1. Claude dynamically generates a JavaScript orchestration script from the prompt
2. It decomposes the task into subtasks and distributes them across dozens or even hundreds of subagents running in parallel
3. One batch of subagents attacks the problem from different angles while another batch rebuts the first batch's findings
4. The whole process iterates until the results converge
5. Everything merges into a unified output handed back to the user

This differs fundamentally from Claude Code's earlier subagent mechanism: previously, intermediate results returned to the conversation context and consumed tokens, whereas Dynamic Workflows moves the orchestration logic into a code script, keeping only the final result in Claude's context.

Flagship case: porting Bun from Zig to Rust

Bun founder Jarred Sumner used Dynamic Workflows to port the JavaScript runtime Bun from Zig to Rust:

• One workflow mapped the correct Rust lifetime for every struct field
• The next workflow wrote a behavior-matching .rs version for every .zig file
• Hundreds of agents worked in parallel
• First commit to merge took 11 days, producing roughly 750,000 lines of Rust code
• 99.8% of the existing test suite passes

How to trigger it

Two ways:

• Option 1: in Claude Code, just say "create a dynamic workflow..." (with the keyword "workflow" in the prompt)
• Option 2: enable the Ultracode setting and Claude decides on its own when to use workflows

Note: Dynamic Workflows burns noticeably more tokens than an ordinary Claude Code session. On first trigger, Claude Code shows you what it is about to run and asks you to confirm.

Pricing and Model Details

Who Should Use It

• Professional developers: Opus 4.8's precision and honesty are greatly improved; pair it with Dynamic Workflows for large-scale code migrations
• Team collaboration: Dynamic Workflows suits bulk edits and defect hunts across hundreds of files
• Long-running tasks: Opus 4.8 can work for extended stretches without a human checking back in frequently

Things to Watch

• Dynamic Workflows is currently a research preview and still iterating
• If you rely on the AI's initiative ("guess what you meant"), adjust your interaction style and state requirements more explicitly
• Creative writing has improved over Opus 4.7, but still trails Opus 4.6