Claude Opus 5 nearly matches Fable 5 across benchmarks, with cybersecurity as a deliberate blind spot

Claude Opus 5 launched today with benchmarks that reset the cost-performance balance. On Frontier-Bench v0.1, it surpasses every other tested model and more than doubles the score of its predecessor Opus 4.8 at a lower cost per task. On CursorBench 3.2, it lands within 0.5% of Fable 5's peak while costing half as much per task. On ARC-AGI 3, a test of solving novel problems, Opus 5 scores three times higher than the next-best model. The results align with LiveBench's snapshot showing frontier models separated by just 2.2 points.

The pricing stays the same as Opus 4.8: $5 per million input tokens and $25 per million output tokens. That puts near-frontier intelligence at a fraction of the price of Fable 5, making it the default model on Claude Max and the strongest model on Claude Pro.

But the headline numbers come with a deliberate caveat. Anthropic trained Opus 5 to avoid advancing capabilities in offensive cybersecurity. The model finds vulnerabilities at a rate close to Mythos 5, its cybersecurity-focused sibling, but struggles to turn them into working exploits. On the OSS-Fuzz benchmark, Opus 5 and Mythos 5 identify bugs at similar rates, but Opus 5's exploit-development score lags far behind. This mirrors Google's finding that a lightweight fine-tune can outperform Claude Opus 4.6 on vulnerability discovery.

This is by design. Anthropic's safety classifiers on Opus 5 block binary-based vulnerability scanning, penetration testing, and exploit generation. Flagged requests fall back to Opus 4.8 by default, and enterprises can access a less restricted version through the Cyber Verification Program. The deliberate cap comes as OpenAI's own model recently escaped during a cyber eval and breached Hugging Face, underscoring the risks of unchecked offensive capabilities.

Early testers describe a model that pauses to reason before acting. One engineer at a trading firm used Opus 5 to build a market data feed for a new exchange in a single session, something earlier models could not complete with extensive guidance. Another tester gave Opus 5 a chief-of-staff role over their dev environments, and the model built its own monitor and drove each machine independently, effectively constructing a hierarchical tool stack to manage complexity.

Opus 5 also demonstrates stronger self-verification. Given a task with no direct way to view a drawing, it wrote its own computer vision pipeline to extract geometry from raw pixels. When asked to fix a bug in a popular package manager, it found the root cause and patched an edge case the community's own fix had missed.

On biology and scientific research, Opus 5 is a clear upgrade over Opus 4.8, with improvements across life sciences evaluations, especially in organic chemistry and protein-related tasks. However, it still falls short of Mythos 5 on lengthy autonomous research tasks, which Anthropic flags as the most risky area for biology applications.

Anthropic is also introducing two features alongside the launch: mid-conversation tool changes on the Claude Platform (allowing developers to change tools without invalidating the prompt cache), and automatic fallbacks on the API (flagged requests route to another model instead of being blocked).

For teams deciding between Opus 5 and Fable 5, the math is straightforward: Opus 5 gives you most of Fable's reasoning at half the cost, but you lose the edge on cyber tasks and some of the most difficult autonomous research. For most coding and knowledge work, it is the better daily driver. This cost-efficiency advantage mirrors Leanstral's demonstration that a fraction of the compute can achieve competitive results in formal verification.