In brief

  • Released on July 24, Claude Opus 5 is priced at $5 per million input tokens, matching its predecessor, Opus 4.8, and is exactly 50% cheaper than Fable 5, while outperforming it across most significant benchmarks.
  • On the Frontier-Bench v0.1, which evaluates coding capabilities, Opus 5 achieved a score of 43.3%, outpacing Fable 5's 33.7% and OpenAI's GPT-5.6 Sol's 34.4%. In the ARC-AGI-3 benchmark, focusing on novel problem-solving, it scored 30.2%, compared to GPT-5.6 Sol's 7.8%—a substantial lead.
  • This new model has become the default on Claude Max and is the most powerful option available on Claude Pro, effectively taking the place of Fable 5 for most users.

Claude Opus 5 is available now. It offers a more economical solution for businesses compared to Anthropic’s flagship model, Claude Fable 5, which was previously promoted as the primary product for paying users. Moreover, Opus 5 demonstrates superior performance on crucial benchmarks.

To clarify the positioning of Opus 5: Anthropic categorizes its models into four tiers. Haiku is the quick and inexpensive option, Sonnet is mid-tier, Opus serves as the robust workhorse, and the Mythos class—introduced this spring—includes Claude Fable 5 for public use and Claude Mythos 5, which is reserved for vetted cybersecurity professionals and critical infrastructure operators through Project Glasswing.

Fable 5 has faced challenges since its launch on June 9; it was withdrawn globally three days later due to a U.S. government emergency export control order that cited a jailbreak vulnerability, and it returned on June 30, but transitioned to a credits-only model, no longer available in standard plans. Opus 5 now occupies the role that Fable 5 could not maintain.

Lovable, a developer platform with millions of users, evaluated Opus 5 internally and found that its improvements go beyond mere scores: "It’s not only better on our toughest coding challenges, showing a 22% increase over Opus 4.7, but it’s also more consistent, with significantly less variation from run to run," said Fabian Hedin in a statement released by Anthropic.

The benchmarks

It may seem unexpected, but Opus outperforms Fable in nearly all areas critical to everyday users, despite not being classified as Mythos-class like Fable.

In the Frontier-Bench v0.1, which assesses whether AI coding agents can complete software engineering tasks from start to finish, Opus 5 scored 43.3%. In comparison, Fable 5 scored 33.7%, while OpenAI’s GPT-5.6 Sol achieved 34.4%.

The most significant difference appeared in the ARC-AGI-3 test, which evaluates genuine problem-solving abilities through unique puzzles that a model cannot memorize from training data. Opus 5 scored 30.2%, while GPT-5.6 Sol scored 7.8%, and Fable 5 was not tested. On GDPval-AA v2—a knowledge work benchmark using Elo ratings, a chess-like ranking system for assessing performance on real tasks—Opus 5 achieved 1,861, compared to Fable 5's 1,747 and GPT-5.6 Sol's 1,736.

Zapier assessed Opus 5 using AutomationBench, which measures a model's ability to manage an entire business workflow autonomously. Their conclusion was that the model "successfully took a raw account-health workbook and executed a complete churn-prevention sequence from start to finish: identifying at-risk accounts, notifying the appropriate owner, and summarizing for retention operations. Previous models did not succeed; Opus 5 achieved 100%."

Anthropic is also promoting Opus 5 as an upgrade for research purposes. Ultima Genomics, a DNA sequencing firm, remarked that the model "acts more like a meticulous scientist than any model we've previously used. It selects appropriate statistical tests to eliminate confounding factors, verifies its own results using independent methods, and maintains focus during extensive multi-step analyses."

However, Fable 5 remains slightly superior in two specific areas: legal and health applications.

This launch follows the recent introduction of Kimi K3 by Moonshot AI, a Beijing-based startup supported by Alibaba, which claims to be the largest open AI system globally with 2.8 trillion parameters (meaning anyone can access the underlying code). Independent assessments consistently rank Kimi K3 third overall, trailing both Fable 5 and GPT-5.6 Sol, but excelling in certain categories.

Opus 5 is currently accessible via API at a cost of $5 per million input tokens and $25 per million output tokens (tokens represent the fundamental unit of information processed by AI models in both input and output). A Fast mode, operating at approximately 2.5 times the standard speed, is also available for double the base price, at $10 per million input tokens and $50 per million output tokens.

This release may alleviate concerns regarding the unavailability of Fable 5 to the public. Opus is also offered through subscription for all users.