Claude Opus 5 Beats Fable 5 in Benchmarks at Half the Price

Introduction

Anthropic announced the launch of Claude Opus 5, positioning it as a more proactive and cautious large‑language model whose intelligence approaches that of Claude Fable 5 but at roughly 50 % of the price.

Performance Benchmarks

In a suite of programming and knowledge‑work evaluations—including Frontier‑Bench, GDPval‑AA, CursorBench 3.2, ARC‑AGI 3, Zapier AutomationBench, and OSWorld 2.0—Opus 5 set new industry records, falling behind Mythos 5 only on a dedicated cybersecurity task.

Programming Capability

On Frontier‑Bench v0.1, Opus 5 more than doubled the score of its predecessor Opus 4.8 while using lower per‑task token cost. In CursorBench 3.2, the model’s highest‑effort setting trails Fable 5 by only 0.5 % but costs half as much; on the high, xhigh, and max settings it outperforms all other models at equal cost.

Knowledge Work and Problem Solving

ARC‑AGI 3 measures a model’s ability to solve novel problems; Opus 5’s score is three times that of the runner‑up. Zapier AutomationBench shows a 1.5× higher task‑completion rate than the second‑best model at the same per‑task cost, and the model remains the top performer even at the lowest effort level. OSWorld 2.0, a computer‑operation benchmark, records Opus 5 beating every competitor at any cost point and achieving roughly one‑third the cost of Fable 5’s best result.

Scientific Capability

Across all life‑science evaluations—structural biology, organic chemistry, and bioinformatics—Opus 5 outperforms Opus 4.8. In organic‑chemistry spectrum‑to‑structure inference, the model improves by 10.2 percentage points; in protein‑variant impact prediction, it gains 7.7 points. Visual outputs also improve, exemplified by aerodynamic flow visualizations and simplified interactive cell diagrams.

Autonomy and Rigorousness

Opus 5 demonstrates stronger self‑verification. In a Frontier‑Bench task requiring code generation from a mechanical‑part diagram without direct image access, the model built a computer‑vision pipeline to extract geometry and successfully reconstructed the part, whereas competing models failed all five attempts. It also identified and patched a deep‑lying vulnerability in a popular open‑source package manager that competitors only superficially addressed. Additionally, an engineer from a trading firm used Opus 5 to build a market‑data source for a new exchange in a single session, with the model creating a test framework to validate its code despite lacking live data.

Alignment and Safety

Anthropic’s pre‑deployment automated behavior audit rates Opus 5 as the most aligned model to date, surpassing Opus 4.8, Sonnet 5, and Fable 5 in adherence to the Claude Charter, with the lowest deception rate and highest resistance to inducement. In the OSS‑Fuzz security evaluation, Opus 5’s vulnerability‑identification success matches Mythos 5, but its ability to develop exploit code lags significantly. Its network‑security classifier is less restrictive than Fable 5’s, allowing source‑code vulnerability scans while blocking binary‑level exploit generation, reducing classifier interventions by roughly 85 %.

Pricing and New Features

Opus 5 is available on all platforms at $5 per million input tokens and $25 per million output tokens—half the cost of Fable 5 and slightly below OpenAI’s GPT‑5 pricing. A new “Fast mode” offers ~2.5× the default speed for double the price in a research‑preview phase. Users can enable automatic fallback, routing requests intercepted by safety classifiers to the next best model. The Claude platform now supports mid‑conversation tool switching without losing prompt cache, and the API adds the same automatic‑fallback capability. Like previous Opus releases, Opus 5 imposes no data‑retention requirement for general access.