Claude Opus 4.8: what’s new, why it matters, and how to use it

Claude Opus 4.8 is here, and it’s a much bigger upgrade than the official announcement suggests. It’s faster, smarter, cheaper in key modes, and introduces new features that can turn months of coding work into an afternoon project—if you know how to use it well.

Claude Opus 4.8 at a glance

Opus 4.8 is Anthropic’s new flagship model, and it currently sits at the top of most major benchmarks. It outperforms GPT-4.5/“5.5”-class models and other competitors across reasoning and coding tests, with Google’s latest models not really in the same range according to early comparisons.

What makes this release especially interesting is that Anthropic has packed in a lot of improvements without raising the base price compared to Opus 4.7. That’s unusual in a landscape where every new model generation has typically come with a cost bump.

Same price, more power

One of the biggest surprises is that Opus 4.8 keeps the same pricing as Opus 4.7. For users who rely on Claude heavily, this is effectively a free performance upgrade.

Behind the scenes, Anthropic has recently gained access to a huge amount of additional compute capacity, which likely explains how they can offer more capability without charging more. The result: you get a stronger model with better reasoning and coding skills for the same money.

Fast mode just became usable

Previously, Claude’s fast mode was so expensive that most people simply ignored it. It cost around six times more than regular mode, which made it hard to justify for everyday use.

In Opus 4.8, fast mode is now three times cheaper than before. That means it’s only about twice as expensive as regular mode, which suddenly makes it a realistic option—especially if you’re on the $200/month plan and need speed for heavier workloads.

If you’ve been leaning on other tools like GPT-based Codex mainly because of their cheap fast modes, this change makes Claude’s fast mode competitive again for coding and iteration-heavy work.

Mythos-level capabilities and fewer hallucinations

Anthropic has been teasing an internal model class called “Mythos” for months—described as extremely capable, especially in security and complex reasoning. Opus 4.8 looks like a slightly toned-down version of that technology, with one standout improvement: hallucinations are reduced by around 4x compared to previous Opus releases.

That reduction lines up with what Anthropic has claimed for Mythos-style models. In practice, this means Opus 4.8 is much less likely to confidently invent facts, APIs, or file structures—especially important when you’re shipping production code or relying on it for research and analysis.

Anthropic has also hinted that “Mythos-class” models will roll out to all customers in the coming weeks, suggesting that Opus 4.8 is a stepping stone toward even more powerful releases.

Dynamic workflows: thousands of sub-agents on demand

One of the most important new features in Opus 4.8 is dynamic workflows. Instead of treating a big request as a single, linear task, Claude can now spin up tens to thousands of sub-agents behind the scenes to tackle different parts of the problem in parallel.

Imagine you ask it to:

• Implement a large new feature across a complex codebase
• One-shot an entire app from scratch
• Refactor and test multiple services at once

Previously, Claude would essentially behave like one agent: read some code, write some code, test a bit, repeat. With dynamic workflows, it can fan out into many workers that simultaneously:

• Touch different parts of your codebase
• Run tests and regression checks
• Explore alternative implementations
• Perform research and validation

This is what enables the “months of work in a day” claim. It’s not magic, but for well-scoped tasks with clear goals, the parallelism can be a huge productivity multiplier. If you’re interested in how Claude handles complex reasoning already, it’s worth comparing this with earlier versions like those discussed in our look at Claude Opus 4.6.

Ultra Code: giving Claude full control

Ultra Code is essentially Claude Code with the safety rails loosened and dynamic workflows fully unleashed. You’re telling Opus 4.8: “Use as much of your internal orchestration as you need to get this done.”

Key points about Ultra Code:

• It’s mainly intended for users on the $200/month plan or higher.
• It’s best used on big, meaty coding tasks—new features, app scaffolding, large refactors—not quick one-liners.
• It can be expensive if you’re on a tight quota, because those sub-agents and extra calls add up.

If you have extra usage enabled (i.e., you’re comfortable paying through the API once you hit plan limits), Ultra Code can become your default for serious builds. Otherwise, you’ll want to reserve it for your most important or complex tasks.

Recommended settings and usage strategy

To get the most out of Opus 4.8 without burning through your limits, here’s a practical setup:

1. Switch all general work to Opus 4.8

There’s no real reason to stay on Opus 4.7. For everyday reasoning, writing, and coding, just select Opus 4.8 as your default model.

2. Be selective with context size

Opus 4.8 supports a 1M-token context window, but filling it up can degrade performance. For most users:

• Use the regular context option by default.
• Only move to the million-token context when you truly need to load huge codebases or long documents.

3. Use “high” effort by default

Claude Code lets you choose effort levels like High, Extra, and Max. A sensible baseline is:

• Default: High
• Upgrade to Extra or Max only for large or critical tasks (big features, major refactors, complex debugging).

Anthropic still doesn’t have quite the same raw capacity as some GPT-based services, so this helps you stay within limits while still getting strong results.

When to use fast mode and Ultra Code

Fast mode and Ultra Code are powerful, but they’re not meant for every single prompt—especially if you’re on a strict budget.

• If you’re on the $200/month plan without extra usage: use fast mode and Ultra Code sparingly, for tasks where speed and depth really matter.
• If you have extra usage enabled and are comfortable paying via API after hitting limits: you can lean much more heavily on fast mode and Ultra Code for daily work.

By contrast, some GPT-based tools give you so much fast-mode capacity that it makes sense to use it for everything. Claude isn’t quite there yet on limits, so a bit of discipline goes a long way.

Don’t switch your agents too early

If you’re using agent frameworks like Hermes or OpenClaw, it’s tempting to flip everything to Opus 4.8 immediately. That often backfires.

New model releases can introduce subtle behavior changes that break assumptions in existing agents, causing crashes or weird edge cases. A safer approach:

• Wait for the official Hermes/OpenClaw releases that explicitly support Opus 4.8 (these usually land within about 24 hours).
• Then migrate your agents once the ecosystem has caught up.

This keeps your automations more stable and saves you from chasing obscure bugs.

Remote control: take Claude Code on the go

Claude Code has a lesser-known “remote control” setting that’s worth turning on. When enabled, every new desktop chat automatically appears on your phone in the Claude app under the Code section.

This lets you:

• Start a coding session on your desktop
• Step away from your desk
• Continue monitoring or nudging the same session from your phone

For anyone who likes to “vibe code” or iterate on ideas away from their main machine, this is a simple but powerful quality-of-life feature.

Real-world coding example: 3D shooter benchmark

To stress-test Opus 4.8’s creativity and coding skills, one benchmark involves asking it to build a 3D first-person shooter using Three.js, with complete creative freedom: power-ups, waves, effects, and whatever else it wants to add.

Compared to earlier models like Opus 4.7, Opus 4.8 produces noticeably better results:

• More polished visuals (walls, floors, lighting, and effects)
• More satisfying shooting mechanics and hit markers
• Cleaner enemy and power-up behavior
• Wave-based gameplay that actually feels like a tiny, coherent game

It’s not going to rival a AAA title, but relative to what previous models generated from the same prompt, it’s a clear upgrade. If you’re curious about how these kinds of coding benchmarks compare across models and workflows, you may also like this guide to being productive with Claude Code and Codex.

Focus is the real multiplier

There’s one non-technical lesson that matters just as much as all the new features: focus. Many people send a prompt to their AI, then immediately drift off into doomscrolling while the model finishes the task—and then they don’t act on the result for another 30–60 minutes.

To actually benefit from tools like Opus 4.8:

• Treat AI as a partner in a flow state, not a background process.
• Put your phone away while you’re working with it.
• Close social media and keep your editor and Claude front and center.
• Respond quickly to its output: test, iterate, refine.

The people who will get the most out of models like Opus 4.8 over the next few years aren’t just the ones with access—they’re the ones who can stay locked in long enough to turn AI output into real shipped products.

Why you should start using Opus 4.8 now

New model releases always create a short window where early adopters have a real advantage. Most people will keep using their old defaults for weeks or months, while a smaller group starts building with the latest tools immediately.

By jumping onto Opus 4.8 now, you can:

• Leverage dynamic workflows and Ultra Code before they’re widely adopted
• Ship more ambitious features and apps faster than your peers
• Learn the model’s strengths and quirks early, so you’re ahead of the curve when Mythos-class models arrive

Block off some focused time, enable Opus 4.8, experiment with fast mode and Ultra Code on a serious project, and see how far you can push it. The tech is ready—the question is whether you are.