Claude Opus 5 leads AI leaderboard at half Fable 5's price

On 24 July 2026, Anthropic announced Claude Opus 5, a new large language model that immediately assumed the top position on the Artificial Analysis leaderboard, surpassing even Claude Fable 5. The model is priced identically to its predecessor Opus 4.8 and maintains the same pricing structure, offering a "fast mode" option at twice the cost of the base model.

Anthropic's own framing emphasizes Opus 5 as "a thoughtful and proactive model that comes close to the frontier intelligence of Claude Fable 5 at half the price." The proactive aspect is demonstrated by a striking example from the Frontier-Bench evaluation suite. In one task, Opus 5 was presented with a drawing of a machine part and asked to write code to rebuild it as a 3D FreeCAD model. Critically, the model was deliberately given no direct way to view the drawing. Rather than failing or requesting clarification, Opus 5 independently wrote its own computer vision pipeline to extract the geometry from the raw pixel data, then used that pipeline to reconstruct the full machine part—solving both the primary task and an unanticipated prerequisite task.

On the cybersecurity front, Anthropic reports that Opus 5 has not been intentionally trained on cyber exploitation tasks, continuing the company's prior approach with Opus 4.8. Nevertheless, the model's general capability improvements have yielded substantial gains in vulnerability detection; it now approaches the performance of Mythos 5 in finding cybersecurity vulnerabilities. However, it remains substantially behind Mythos 5 in the exploitation phase—the ability to turn discovered vulnerabilities into active cyber threats. Anthropic attributes this gap to the deliberate absence of exploitation training and notes that this design choice may prove favorable in discussions with government regulators.

To support developer adoption, Anthropic published a dedicated prompting guide for Claude Opus 5. Additional resources include a guide titled "The new rules of context engineering for Claude 5 generation models," authored by Thariq Shihipar, to help users optimize their interactions with the new model generation.

Claude Opus 5 represents a milestone in Anthropic's model hierarchy: it achieves top ranking on a major independent benchmark while undercutting the cost of the previous generation's flagship model. The pricing structure mirrors Opus 4.8, with a base tier and a "fast mode" available at double the cost, giving users a choice between speed and cost efficiency.

A notable aspect of Opus 5's capabilities is its apparent tendency toward proactive problem-solving beyond what it was explicitly instructed to do. The Frontier-Bench anecdote reveals this clearly: faced with incomplete information (a drawing it could not directly view), the model spontaneously authored a computer vision pipeline as an intermediary step. This suggests that general capability improvements have also expanded the model's ability to decompose and solve novel sub-problems.

Anthropichas deliberately maintained its prior stance on cybersecurity training: Opus 5 was not trained on exploitation tasks, despite improving substantially on vulnerability detection as a side effect of becoming more generally capable. The model now approaches Mythos 5 in vulnerability finding but remains far behind on turning those findings into actual cyber threats—a safety-conscious design choice that the author notes may help with regulatory acceptance.

No comments yet. Be the first to share your thoughts!