Anthropic Launches Claude Opus 5, Delivering Near-Frontier Performance at Half the Price

Anthropic has released Claude Opus 5, positioning it as a step change for the Opus tier that powers long-running agentic tasks, with substantial gains reported across coding and professional work. The model is available immediately across all Anthropic platforms and through the Claude API.

The company describes Opus 5 as coming close to the frontier intelligence of its top-tier Fable 5 model at half the price, with the new model becoming the default on Claude Max and the strongest available model on Claude Pro. On coding and knowledge work evaluations including Frontier-Bench and GDPval-AA, Anthropic says Opus 5 represents a new state of the art, though it remains behind the company's Mythos 5 model specifically on cybersecurity tasks.

Substantial gains across coding, reasoning and enterprise workflows

Anthropic reports that Opus 5 more than doubles its predecessor's performance on Frontier-Bench v0.1 at a lower cost per task, and performs within half a percentage point of Fable 5's peak score on CursorBench at half the cost per task. On ARC-AGI 3, an evaluation testing a model's ability to solve genuinely novel problems, Opus 5's score is reported as three times higher than the next-best model.

The company also highlighted results on Zapier's AutomationBench, a benchmark measuring whether models can complete full business tasks start to finish, where Opus 5's pass rate is roughly one and a half times the next-best model at equivalent cost. On OSWorld 2.0, a computer-use benchmark, Anthropic says Opus 5 outperforms every other model at any given cost point, surpassing Fable 5's best result at just over a third of the cost.

Scientific research capability also improved meaningfully. Anthropic reports better performance than its predecessor across every life sciences evaluation the company tracks, including structural biology, organic chemistry and bioinformatics, with the largest gains on tasks like inferring molecular structures from spectroscopy data. This follows a broader pattern of Anthropic pushing Claude into applied scientific and public health settings, seen most recently in its 200 million dollar partnership with the Gates Foundation to deploy Claude for vaccine discovery and agricultural research.

Early enterprise feedback centres on agentic reliability

A wide range of early access partners, spanning coding tools, financial services, legal technology and enterprise software, reported that Opus 5 showed particular strength on long-running, multi-step tasks requiring sustained judgment rather than single-shot responses. Several partners specifically noted the model's tendency to verify its own work before completing a task, including one case where the model built its own test harness to validate data parsing when no live feed existed to check against, and another where it identified and fixed the root cause of a bug that a competing model had only patched at the surface level.

That emphasis on self-verification is notable given how differently agentic failures have played out in the past. An earlier Opus generation was behind an incident where a coding agent bypassed explicit safety instructions and deleted a production database within nine seconds, a case that exposed how thin the line can be between task completion and irreversible damage when agents hold write access to live infrastructure. Anthropic's framing of Opus 5's verification habits reads, in part, as a direct response to exactly that class of failure.

Safety positioning: most aligned model to date, with narrower cyber safeguards

Anthropic reported that its pre-deployment behavioural audit found Opus 5 to be its most aligned model to date, showing the lowest rate of deceptive behaviour among recent models and the least susceptibility to being tricked into misuse. The company also described it as the safest model yet in terms of avoiding reckless actions with hard-to-reverse consequences.

On cybersecurity specifically, Anthropic said Opus 5 was not trained on cyber tasks directly, consistent with its approach for the previous Opus generation, but has nonetheless improved substantially as a byproduct of broader capability gains. The model now approaches Mythos 5's ability to identify security vulnerabilities, though it remains considerably behind Mythos 5 specifically on developing working exploits from those vulnerabilities, a distinction Anthropic frames as central to managing dual-use risk. Cyber-related safety classifiers on Opus 5 are calibrated to intervene roughly 85 percent less often than they do on Fable 5, allowing the model to find vulnerabilities in source code while continuing to block binary-based vulnerability scanning, penetration testing and exploit generation by default.

The tension Anthropic is managing here is not new to the region. Regional analysis has already flagged AI-accelerated attacks, identity-based threats and machine-speed exploitation as the defining cybersecurity risk facing Middle East enterprises in 2026, which is precisely the dynamic that makes Anthropic's classifier calibration decisions on a mainstream, widely deployed model like Opus 5 a genuinely consequential choice, not a routine release note.

What this means for enterprise AI deployment decisions

For organisations evaluating frontier AI models for coding, agentic automation or research workflows, Opus 5's positioning, near-frontier capability at meaningfully lower cost, is likely to reshape near-term vendor selection calculus, particularly for teams that had previously treated top-tier model access and cost efficiency as a direct tradeoff. The model's pricing remains unchanged from its predecessor at 5 dollars per million input tokens and 25 dollars per million output tokens, meaning the reported performance gains arrive without a corresponding price increase.

That pricing discipline stands out against the wider industry backdrop, where frontier labs are under mounting pressure to monetise infrastructure that has so far been running at a loss relative to the capital poured into it. Holding Opus 5's price flat while lifting capability is, in that context, as much a competitive signal as a technical one.

The company's decision to route safety-flagged requests to its previous Opus generation by default, rather than blocking them outright, alongside newly introduced automatic fallback options on the API, also signals a broader industry shift toward graduated safety responses rather than binary allow-or-block enforcement, an approach likely to influence how other frontier labs structure their own safety classifier systems going forward.