Claude Opus 5 Launches: Matches Frontier AI at Half Cost, Still Trails in Cyber Exploits

Anthropic released Claude Opus 5 today, positioning it as a model that closes much of the gap with its flagship Claude Fable 5 while cutting inference costs roughly in half.

The launch signals a shift in AI provider strategy: rather than chasing raw capability alone, Anthropic is betting that cost-efficiency and daily usability will drive enterprise adoption, particularly for coding and knowledge-work workflows.

Opus 5 sets new state-of-the-art marks on Frontier-Bench v0.1 and GDPval-AA, more than doubling predecessor Opus 4.8’s performance on software engineering tasks at a lower per-task cost. On CursorBench 3.2, running at max effort, it lands within 0.5% of Fable 5’s peak score while costing half as much per task.

Claude Opus 5 Launches

The model also posted strong results on:

ARC-AGI 3 (novel problem-solving): three times the score of the next-best model
Zapier AutomationBench (end-to-end business tasks): roughly 1.5x the pass rate of competitors at equivalent cost
OSWorld 2.0 (computer-use benchmark): beats every rival model at any cost tier, surpassing Fable 5’s best result at just over a third of the price

Anthropic also highlighted gains in life sciences applications, including a 10.2 percentage point jump over Opus 4.8 on organic chemistry tasks like molecular structure inference from spectroscopy data.

For security practitioners, the most consequential detail sits in the model card, not the headline. Anthropic explicitly states Opus 5 “remains behind Mythos 5 on cybersecurity tasks” and was intentionally not trained on offensive cyber tasks.

The company’s OSS-Fuzz evaluation, designed to test vulnerability discovery and exploit development without human guidance, shows Opus 5 nearly matching Mythos 5 at identifying vulnerabilities but falling substantially behind on turning those vulnerabilities into working exploits. This vulnerability-discovery-versus-exploitation gap is Anthropic’s clearest attempt yet to quantify dual-use risk in a shipped model.

Opus 5’s cyber classifiers are described as “proportionally less restrictive” than Fable 5’s, intervening roughly 85% less often. The model can hunt for vulnerabilities in source code, but binary-based vulnerability scanning, penetration testing, and exploit generation remain blocked by default. Flagged requests fall back automatically to Opus 4.8 across Claude.ai, Claude Code, and Claude Cowork.

Organizations enrolled in Anthropic’s Cyber Verification Program (CVP) get access to a less-restricted variant, a notable carve-out that effectively creates a two-tier access model for security research versus general use.

Anthropic’s automated behavioral audit scored Opus 5 at 2.3 on overall misaligned behavior, its lowest (best) figure to date, with reduced deceptive behavior and lower susceptibility to misuse or manipulation compared to Opus 4.8, Sonnet 5, and Fable 5.

Pricing holds steady at $5 per million input tokens and $25 per million output tokens, matching Opus 4.8 despite the performance jump. A Fast mode variant runs at 2.5x default speed for double the base price. The model is now the default on Claude Max and the strongest option available on Claude Pro, available immediately via the Claude API as claude-opus-5.

For threat intelligence teams, the takeaway is nuanced: Opus 5 narrows the capability gap in vulnerability discovery, but Anthropic’s own data suggests it isn’t yet the exploit-generation risk that Mythos 5 represents, at least not without CVP-level access.