AI News (2026/7/26): Anthropic Launches Claude Opus 5: Near-Flagship Intelligence at Half the Price

2026年7月26日 01:20

Executive Summary:

On July 24, Anthropic released the flagship model Claude Opus 5, positioned as "a cutting-edge intelligence close to Fable 5, at approximately half the cost." Compared to the previous generation, Opus 4.8, the company claims a significant performance boost without increasing cost...

Model Overview and Positioning

On July 24, Anthropic released the flagship model Claude Opus 5, positioned as "a cutting-edge intelligence close to Fable 5, at approximately half the cost." Compared to the previous generation, Opus 4.8, the company claims a significant performance boost without increasing costs. Opus 5 introduces adjustable effort thinking modes (low, medium, high, xhigh, max), allowing users to balance between stronger reasoning capabilities, lower latency, and reduced token usage. This model is now the default for Claude Max and also the highest capability option available on Claude Pro. The company describes it as a more proactive and thoughtful model, capable of reducing back-and-forth confirmations, self-checking, and recovering from errors during tasks, while emphasizing its safety guardrails.

Claude Opus 5 - Anthropic Official Website Screenshot
Image source: official article

Core Capabilities and Benchmark Performance

According to Anthropic's benchmark data, Opus 5 leads in multiple evaluations:

  • Terminal Programming: Achieved a score of 43.3% on Frontier-Bench v0.1, surpassing Fable 5 (33.7%), GPT-5.6 Sol (34.4%), and Opus 4.8 (21.1%). The official claims that its performance is more than double that of Opus 4.8, with lower single-task costs.
  • Knowledge Work: Scored 1861 on GDPval-AA v2, outperforming Fable 5 (1747), GPT-5.6 Sol (1736), and Opus 4.8.
  • Novel Problem Solving: Scored 30.2% on ARC-AGI 3, approximately three times that of the second-place GPT-5.6 Sol (7.8%).
  • Business Process Automation: Passed 26.0% of AutomationBench tests, achieving about 1.5 times the performance of the second-best model at the same cost.
  • Computer Operation: Scored 70.6% on OSWorld 2.0, higher than Fable 5 (66.1%) and GPT-5.6 Sol (62.6%). The official states that it can achieve the best performance of Fable 5 at about one-third of the cost.
  • Intelligent Search: Scored 90.8% on BrowseComp, slightly higher than GPT-5.6 Sol (90.4%).
  • Scientific Research: Outperformed Opus 4.8 comprehensively in life sciences evaluations, with improvements of 10.2 percentage points in spectral inference of molecular structures and 7.7 percentage points in predicting the functional impact of protein sequence variations.

In addition, Opus 5 is only 0.5% behind Fable 5 in the highest tier of CursorBench 3.2, with about half the cost. Its cybersecurity capabilities have also seen a significant overall improvement compared to Opus 4.8. On software engineering benchmarks such as DeepSWE v1.1 and FrontierCode v1.1, Opus 5's scores are close to those of Fable 5.

Usage and Subscription

Claude Opus 5 is now available through the Claude application (web and mobile versions). Claude Max subscribers automatically use this model, while Claude Pro users can select it as the highest capability option. The interface provides entry points such as "Opus 5 High" for different effort levels, allowing users to choose the thinking intensity based on task complexity. In terms of API, Opus 5 offers a 1 million token context window and a maximum output of 128,000 tokens, which can be accessed via the Anthropic API and cloud platforms such as AWS (Amazon Bedrock) and Microsoft Foundry. The official has not yet disclosed details such as the model's parameter count and the composition of the training data.

Sources and Project Links

Related AI Tools

About AI News

We use AI technology to automatically crawl and filter the latest AI news from around the world, providing you with the most valuable industry updates.