Claude Opus 4.8: Features, Benchmarks & Complete Guide

Claude Opus 4.8 is Anthropic's new flagship model, released on May 28, 2025. As the direct successor to Opus 4.7, it stands out through increased honesty in its responses, the ability to orchestrate hundreds of sub-agents in parallel, and improved performance on agentic coding benchmarks. Available at the same price as Opus 4.7 — $5 per million input tokens and $25 per million output tokens — it targets developers, product teams, and anyone looking for a more reliable AI assistant.

What is Claude Opus 4.8?

Claude Opus 4.8 is a large language model (LLM) developed by Anthropic. It carries the API identifier claude-opus-4-8 and replaces claude-opus-4-7 in the Claude model family hierarchy. Unlike a full generational leap (as Claude Mythos will be), Opus 4.8 represents a targeted iteration: Anthropic focused its efforts on three axes — honesty, agentic efficiency, and generated code quality — rather than expanding the context window or overhauling the architecture.

The model is accessible through:

• The Anthropic API (endpoint claude-opus-4-8)
• Claude.ai (Pro subscription at $20/month, Team, Enterprise, Max)
• Claude Code (terminal)
• Claude Cowork (real-time collaboration)

Anthropic describes Opus 4.8 as "a more effective collaborator": a model that no longer merely produces text but knows how to recognize its limitations, flag its uncertainties, and adjust its strategy along the way.

Honesty as a core feature

The word appears in every official Anthropic communication about Opus 4.8: honesty. The concept covers several concrete behaviors that the model now adopts more systematically.

Fewer unsupported claims

Opus 4.8 is trained to reduce factual hallucinations. When it lacks a reliable source to answer a question, it says so explicitly instead of inventing a plausible response. This change is particularly noticeable in technical domains (law, medicine, engineering) where incorrect information can have real consequences.

Proactive bug reporting

According to Anthropic, Opus 4.8 is approximately four times less likely than its predecessor to let a code defect pass without flagging it. In practice, this means that when you ask Claude to generate or review code, it will warn you about potential flaws rather than silently ignoring them.

Let's take a concrete example. If you ask Opus 4.7 to create an API route without input validation, it frequently produces the requested code without mentioning the risk. Opus 4.8, on the other hand, will add a warning like:

Questioning its own plans

Tom Pritchard, a staff engineer at Shopify and early adopter of Opus 4.8, summarizes this behavior well: the model "asks the right questions, catches its own mistakes, pushes back when a plan isn't sound." It is no longer a simple executor; it's a counterpart that demonstrates critical judgment.

This behavior is particularly useful in Claude Code, where sessions can last for hours. Instead of going down the wrong path for 200 lines of code, Opus 4.8 stops, explains why the initial approach doesn't work, and proposes an alternative.

Benchmarks: how Opus 4.8 stacks up

The benchmarks published by Anthropic show significant improvements across several axes. Here is a summary comparison:

Opus 4.8's strengths

The model clearly dominates on SWE-Bench Pro, a benchmark that measures an AI agent's ability to solve real GitHub issues end-to-end. At 69.2%, Opus 4.8 widens the gap with GPT-5.5 (58.6%) and confirms Anthropic's lead in agentic coding.

On Online-Mind2Web, which evaluates a model's ability to autonomously navigate a web browser, Opus 4.8 reaches 84%, ahead of its direct competitors. This result is consistent with the model's agentic orientation.

The GDPval-AA score (1,890) is the highest on the board, suggesting that Opus 4.8 excels in professional tasks: writing, data analysis, planning.

Points to consider

On Terminal-Bench 2.1, GPT-5.5 retains the advantage (78.2% vs. 74.6%). This benchmark measures the ability to execute commands in a terminal, a scenario where OpenAI's model remains slightly ahead. Moreover, all these figures come from Anthropic and have not yet been confirmed by independent evaluations. As always, it's wise to wait for community testing before forming a definitive opinion.

Dynamic workflows: hundreds of sub-agents in parallel

The most ambitious feature of Opus 4.8 is called dynamic workflows. Launched as a research preview, it allows Claude Code to plan complex work, spawn hundreds of parallel sub-agents, and verify results before reporting back to the user.

How does it work?

Imagine you need to migrate a 300,000-line codebase from one framework version to another. With traditional tools, you manually divide the work, make changes file by file, and handle conflicts by hand. With dynamic workflows:

1. Planning: Claude analyzes the entire repository, identifies modules to migrate, and creates an action plan.
2. Parallel execution: it launches sub-agents that work simultaneously on different parts of the code.
3. Verification: each sub-agent validates its results (existing tests, consistency) before reporting back.
4. Dynamic adaptation: if a sub-agent encounters an unexpected problem, the overall plan is adjusted in real time.

This last point is crucial. Unlike a fixed-plan system, dynamic workflows allow agents to modify their priorities based on what they discover during execution. A blocking bug in one module can trigger a complete reorganization of the queue.

The connection to honesty

This multi-agent architecture makes the model's honesty even more critical. When Claude orchestrates hundreds of sub-agents, no human can verify each result in real time. The sub-agents must therefore be capable of:

• Flagging their uncertainties instead of masking them
• Detecting errors in their own work
• Reporting failures rather than covering them up

This is precisely what the honesty improvement in Opus 4.8 enables at scale.

Effort control: choosing between speed and depth

Effort control already existed in Claude Code since Opus 4.7, but Opus 4.8 extends it to Claude.ai and Claude Cowork. The principle is simple: a slider lets you define how much "thinking" Claude invests in each response.

The different effort levels

With low effort, Claude responds faster and consumes fewer tokens, reducing cost and throttling. With maximum effort, the model "thinks more frequently and more deeply," in Anthropic's words.

Impact on costs

In low effort mode, you can cut your token consumption by two to three times on certain tasks, representing significant savings for teams using the API at scale. Opus 4.8's default mode consumes a similar number of tokens to Opus 4.7, but with better performance, which amounts to a free improvement in value for money.

Pricing: what changes and what stays the same

Anthropic maintains Opus 4.7 pricing for the standard version of Opus 4.8:

Fast mode becomes three times cheaper

The real pricing news concerns Fast mode. Anthropic announces it now costs three times less than on previous models. This mode multiplies generation speed by 2.5, making it attractive for interactive use cases where latency is a bottleneck.

For consumer users, Opus 4.8 is accessible starting with the Pro plan at $20/month ($17 with annual billing). Team and Enterprise plans offer more generous quotas and access to dynamic workflows.

Comparison with the competition

Opus 4.8 remains significantly more expensive than competing models. OpenAI's GPT-5.5 and Google's Gemini models offer lower pricing for sometimes comparable performance. Anthropic's positioning is clear: Opus is a premium model, and price competition will come from future lighter models offering "Opus-like" capabilities at lower cost.

Using Opus 4.8 via the Anthropic API

Integrating Opus 4.8 into your projects goes through the Anthropic API. Here's how to make a basic call in Python:

import anthropic

client = anthropic.Anthropic(api_key="your-api-key")

response = client.messages.create(
    model="claude-opus-4-8",
    max_tokens=4096,
    messages=[
        {
            "role": "user",
            "content": "Analyze this code and flag any potential issues: ..."
        }
    ]
)

print(response.content[0].text)

Configuring the effort level

To adjust the model's thinking level through the API, you can use the thinking token budget parameter:

response = client.messages.create(
    model="claude-opus-4-8",
    max_tokens=8192,
    thinking={
        "type": "enabled",
        "budget_tokens": 5000  # Adjust based on complexity
    },
    messages=[
        {
            "role": "user",
            "content": "Design a microservices architecture for an e-commerce platform."
        }
    ]
)

A higher thinking token budget corresponds to greater effort. For simple tasks, reduce this budget to save tokens and speed up the response.

Node.js example

If you work with JavaScript or TypeScript, here's the equivalent with the official SDK:

import Anthropic from "@anthropic-ai/sdk";

const client = new Anthropic({ apiKey: "your-api-key" });

const response = await client.messages.create({
  model: "claude-opus-4-8",
  max_tokens: 4096,
  messages: [
    {
      role: "user",
      content: "Review this pull request and list the critical issues."
    }
  ]
});

console.log(response.content[0].text);

Claude Code and Opus 4.8: what changes in practice

Claude Code is Anthropic's terminal environment, designed for developers to interact with Claude directly from their IDE or command line. Opus 4.8 brings several concrete improvements.

Automatic model update

When Anthropic deploys a new model, Claude Code doesn't always detect it immediately. During the transition from Opus 4.6 to Opus 4.7, some users had to wait until the next day for the new model to become active. If you've just started a session and Claude Code is still using Opus 4.7, try restarting your session or waiting a few hours.

You can check the active model by simply asking:

# In Claude Code
> What model are you currently using?

Better judgment in long sessions

Development sessions with Claude Code often last several hours. Opus 4.7 had already improved the model's ability to change strategy when an approach failed. Opus 4.8 takes this logic further: the model doesn't just pivot, it explains why it's doing so, allowing you to follow its reasoning and course-correct if needed.

This transparency is particularly valued on complex projects where context evolves throughout the session. Instead of producing code that "seems to work" but contains hidden flaws, Opus 4.8 prefers to stop and ask you a clarifying question.

Opus 4.8 vs. the competition: GPT-5.5, Gemini 3.5 Flash

The launch of Opus 4.8 doesn't happen in a competitive vacuum. Google just deployed Gemini 3.5 Flash, a model optimized for speed and low cost. OpenAI continues to evolve GPT-5.5, which dominates several generalist benchmarks.

Where Opus 4.8 excels

• Agentic code: best score on SWE-Bench Pro, far ahead of GPT-5.5 and Gemini
• Autonomous navigation: leader on Online-Mind2Web
• Honesty and reliability: drastic reduction in unreported bugs
• Multi-agent orchestration: dynamic workflows have no direct equivalent from competitors

Where competitors maintain the edge

• Pure terminal: GPT-5.5 stays ahead on Terminal-Bench 2.1
• Price: Gemini 3.5 Flash and GPT-5.5 are significantly cheaper
• Ecosystem: OpenAI benefits from a larger user base and tool ecosystem

The choice between these models depends on your use case. For agentic coding and multi-agent workflows, Opus 4.8 is currently the best choice. For simple, high-volume tasks, a cheaper model will be more cost-effective.

Claude Mythos: what comes after Opus 4.8

Opus 4.8 isn't the end of the road for Anthropic. In the background, the company is working on Claude Mythos, a model of a "superior intelligence class" that sits above the Opus family in the internal hierarchy.

Mythos is already in the hands of a limited number of organizations through the Project Glasswing program, primarily for cybersecurity work. Anthropic indicates that broader availability of Mythos is planned "in the coming weeks," without a specific date.

This positioning confirms Anthropic's strategy: pushing performance boundaries with Mythos while ensuring that safety guardrails are battle-tested on production models like Opus before scaling up.

Best practices to get the most out of Opus 4.8

If you already use Claude or are considering switching to Opus 4.8, here are some recommendations to maximize the model's value.

Leverage the model's honesty

Don't disable Claude's warnings and don't rephrase your prompts to force an unqualified answer. One of Opus 4.8's greatest strengths is its ability to say "I'm not sure." Welcome these moments as valuable information rather than failures.

Adjust effort based on the task

Use low effort for brainstorming, rephrasing, and quick questions. Reserve maximum effort for architecture problems, complex debugging, and critical code reviews. This discipline will save you tokens and reduce latency.

Structure your prompts for long sessions

In Claude Code, provide a context file (like a CLAUDE.md) at the root of your project. This file guides Claude's behavior throughout the session and reduces the risk of drift in long conversations.

# CLAUDE.md - Project context

## Tech stack
- Framework: Next.js 15
- Database: PostgreSQL with Prisma
- Tests: Vitest + Testing Library

## Conventions
- Use Server Components by default
- Name files in kebab-case
- Every component must have a test

## Watch points
- Legacy auth system migration is in progress
- Do not modify files in /legacy without asking

Leverage dynamic workflows for large migrations

If you're on an Enterprise, Team, or Max plan, dynamic workflows are designed for large-scale tasks. Instead of migrating a repository file by file, describe the overall migration and let Claude orchestrate the sub-agents. Make sure your test suite is comprehensive before launching a migration: sub-agents rely on existing tests to validate their changes.

Frequently asked questions

Opus 4.8 primarily improves the model's honesty (four times fewer unreported code bugs), performance on agentic coding benchmarks (69.2% on SWE-Bench Pro vs. 64.3%), and introduces dynamic workflows, which allow launching hundreds of sub-agents in parallel. Pricing remains identical to Opus 4.7.

The model is priced at $5 per million input tokens and $25 per million output tokens, exactly the same rate as Opus 4.7. Fast mode (2.5× faster) is $10 input and $50 output, representing a price three times lower than what fast mode cost on previous generations. Consumer subscriptions start at $20/month (Pro plan).

It depends on the use case. Opus 4.8 outperforms GPT-5.5 on agentic code (SWE-Bench Pro) and autonomous navigation (Online-Mind2Web). However, GPT-5.5 leads on Terminal-Bench 2.1 and offers lower pricing. For agentic development, Opus 4.8 is currently the most performant choice.

Dynamic workflows are a research preview feature that allows Claude Code to plan complex work, launch hundreds of sub-agents in parallel, and verify results before reporting back. Unlike a fixed plan, agents adjust their priorities based on the problems they encounter. This feature is restricted to Enterprise, Team, and Max plans.

Anthropic indicates that "Mythos-class" models will be more broadly available "in the coming weeks," but without giving a specific date. Mythos is currently accessible only to a limited number of organizations through the Project Glasswing program, primarily for cybersecurity work. Anthropic plans to integrate the safeguards developed from Mythos into a future Opus model before a public deployment.