Preface¶

On July 24, 2026, Anthropic officially launched Claude Opus 5, positioning it as the most powerful model in the Opus series to date. The core selling point from the official announcement is straightforward: its intelligence level in coding and knowledge work scenarios is close to that of the flagship Claude Fable 5, but the API pricing remains at the same level as Opus 4.8 — $5 per million input tokens and \(25 per million output tokens, roughly half the cost of Fable 5 (\)10/$50).

In the week following the launch, the official announcement post on Hacker News received over 1,700 upvotes and 1,300+ comments, with developer discussions focusing on Agent coding, long-range tasks, and the cost-performance ratio of “cutting-edge capabilities + Opus pricing.” This article is based on Anthropic’s official announcements, product pages, System Card, and AWS official blog posts to sort out the key changes and access methods for Claude Opus 5.

Launch Overview: Pricing Unchanged, Capabilities Leap Forward¶

Claude Opus 5 is now available across all platforms, including:
- Claude Client: Default model for Claude Max, the most powerful optional model for Claude Pro
- Claude API: Model ID is claude-opus-5
- Claude Code / Claude Cowork: Supports agent coding and collaboration scenarios
- Cloud Vendors: Amazon Bedrock, Google Cloud, Microsoft Foundry

Pricing is exactly the same as Opus 4.8:

Additionally, Fast Mode is available: inference speed is approximately 2.5x the default mode, with double the price (consistent with Opus 4.8’s strategy). General access has no data retention requirements, and enterprises can configure it as needed.

Regarding context window, Anthropic states on the Opus product page and System Card that the Opus series supports a 1 million token context; the System Card notes that the effective context in various evaluation scenarios does not exceed 1 million tokens.

Benchmark Tests: New SOTA for Coding and Knowledge Work¶

Anthropic emphasized in its launch announcement that Opus 5 has reached a new state-of-the-art on coding and knowledge work benchmarks such as Frontier-Bench and GDPval-AA, but lags behind Claude Mythos 5 in cybersecurity tasks.

Some key data disclosed in the System Card and official blog posts are as follows:

Software Engineering¶

The official blog also mentioned that on CursorBench 3.2, the peak gap between Opus 5 and Fable 5 under max effort is within 0.5%, but the per-task cost is about half; on Frontier-Bench v0.1, Opus 5 outperforms all comparison models, with double the performance of Opus 4.8 and lower per-task costs.

Reasoning, Automation, and Computer Use¶

• ARC-AGI 3 (requires interaction to solve novel problems): Opus 5 scores 30.2% under high effort, with the next-best model at approximately 7.8% according to the official statement
• OSWorld 2.0 (computer use): Outperforms Fable 5’s best results at about one-third the cost
• AutomationBench (end-to-end business automation): Pass rate is approximately 1.5x that of the second-best model, exceeding all comparison models even at the lowest effort level
• GDPval-AA v2 (knowledge work Elo): 1861, higher than Opus 4.8’s 1593

Effort Levels: Trading “Inference Strength” for Cost¶

Opus 5 continues the effort adjustment mechanism from the Claude 5 series, allowing users to choose inference depth based on task difficulty via the API and Claude platform. The levels listed in the System Card include low, medium, high, xhigh, max.

Taking Frontier-Bench v0.1 as an example (official data):
- max / xhigh: ~43–44% mean reward
- high: 39% mean reward, with an average ~19% reduction in output tokens
- low: 25% mean reward, with an average ~64% reduction in output tokens

This means lightweight tasks such as daily script repairs and document organization can use lower effort levels to control costs; complex refactoring and cross-module debugging require higher levels such as xhigh / max. Third-party evaluations from tools like Cursor and Devin also note that Opus 5 approaches Fable-level performance on FrontierCode 1.1 at half the cost.

Agent Coding: Validation, Iteration, and Long-Range Tasks¶

Anthropic positions Opus 5 as a “daily usable” agent coding model, highlighting its improvements in the following areas:
1. Self-validation: More inclined to check work before submitting, rather than rushing to output
2. Root cause analysis: In real open-source package manager bug cases, Opus 5 identifies root causes and fixes edge cases missed by community patches
3. Long-range autonomy: Can maintain context across multiple-step tool calls, suitable for agent pipelines that run for hours or even overnight

Feedback from official Early Access customers shows that the Cursor team stated Opus 5 is “just under Fable 5” on CursorBench with similar behavior; JetBrains emphasized its ability to detect logical errors during the planning phase; internal evaluations from Lovable show that agentic coding tasks have improved by approximately 22% compared to Opus 4.7, with significantly reduced run-to-run variance — for production agents, stability is as important as peak scores.

Claude Code users can directly switch to claude-opus-5 within the IDE; Fast Mode can be enabled via usage credits, suitable for latency-sensitive scenarios.

Safety and Alignment¶

The System Card discloses that Opus 5 has a misaligned behavior score of 2.3 in automated behavior audits, the lowest among recent Claude models. In terms of cybersecurity, Opus 5 is close to Mythos 5 in vulnerability identification but lags significantly in exploit development; Mythos 5 remains stronger in high-risk long-range autonomous research tasks related to biology.

Two new Beta capabilities have been added to the API:
- Mid-conversation tool changes: Add or remove tools mid-conversation without invalidating the Prompt Cache
- Automatic fallbacks: Automatically route to other available models (such as Opus 4.8) when blocked by the safety classifier

How to Access Claude Opus 5¶

Claude API (Python)¶

import anthropic

client = anthropic.Anthropic()

message = client.messages.create(
    model="claude-opus-5",
    max_tokens=4096,
    messages=[
        {
            "role": "user",
            "content": "Review this Python module for race conditions and suggest fixes.",
        }
    ],
)

print(message.content[0].text)

Amazon Bedrock (Boto3)¶

The Bedrock invocation example provided in the AWS official blog post:

import boto3
import json

bedrock_runtime = boto3.client(
    service_name="bedrock-runtime",
    region_name="us-east-1",
)

response = bedrock_runtime.invoke_model(
    modelId="global.anthropic.claude-opus-5",
    contentType="application/json",
    accept="application/json",
    body=json.dumps({
        "anthropic_version": "bedrock-2023-05-31",
        "max_tokens": 4096,
        "messages": [
            {
                "role": "user",
                "content": "Design a distributed architecture on AWS in Python "
                           "that should support 100k requests per second "
                           "across multiple geographic regions.",
            }
        ],
    }),
)

result = json.loads(response["body"].read())
print(result["content"][0]["text"])

Bedrock defaults to Zero Data Retention; Claude Platform on AWS can apply for ZDR on demand, and integrates with AWS billing and IAM. The model is now available in multiple regions including US East (N. Virginia), Europe (Ireland), and Asia Pacific (Melbourne).

What the Developer Community Thinks¶

The HN discussion (Claude Opus 5 announcement thread) presents polarized views:

Positive: Cases such as spontaneously building ML pipelines and reconstructing 3D FreeCAD models from pixels demonstrate Opus 5’s “proactivity”; many users believe that near-Fable 5 capabilities paired with Opus pricing are extremely attractive in agent coding scenarios.

Controversy: Some developers feedback that Opus 5 is more “opinionated”, sometimes bypassing user-specified rules or performing unrequested extra work, leading to increased token consumption; in the initial launch period, Anthropic and Bedrock also experienced elevated errors, with some users reporting more stable access via AWS Bedrock than direct API connections.

Combining official data and community feedback, Opus 5 is suitable for the following scenarios:
- Long-range Agentic Coding (Claude Code, Cursor, Devin, etc.)
- Cross-session enterprise knowledge work (spreadsheets, documents, presentations)
- Large codebase analysis requiring a 1 million token context
- Teams looking to strike a balance between Fable-level intelligence and Opus-level pricing

If tasks primarily consist of simple Q&A, or if you are extremely sensitive to latency and stability, Sonnet 5 or Opus 4.8 may still be more reliable choices — the value of Opus 5 lies in bringing “cutting-edge agent capabilities” to Opus pricing, rather than fully replacing Fable 5 or Mythos 5 across all scenarios.

Reference Links¶

• Introducing Claude Opus 5 (Anthropic Official Announcement)
• Claude Opus Product Page
• Claude Opus 5 System Card
• Introducing Claude Opus 5 on AWS (AWS Official Blog Post)