Anthropic announced Claude Opus 5 on July 24, 2026. The company describes Opus 5 as a more thoughtful and proactive model that approaches the frontier intelligence of Claude Fable 5 while operating at roughly half the cost.
Performance across key tasks
Opus 5 is presented as the new state-of-the-art on several software engineering and knowledge-work evaluations. It leads on benchmarks such as Frontier-Bench v0.1 and GDPval-AA, although it remains behind Mythos 5 on cybersecurity-specific tasks.
The model shows particular strength on software engineering tasks:
- Frontier-Bench v0.1: Opus 5 outperforms all other models, more than doubling Opus 4.8’s performance at a lower cost per task.
- CursorBench 3.2: at max effort, Opus 5 performs within 0.5% of Fable 5’s peak score while using about half the cost per task.
Similar gains appear on knowledge-work and problem-solving evaluations:
- ARC-AGI 3: Opus 5 scores three times higher than the next-best model on this novel-problem-solving evaluation.
- Zapier AutomationBench: Opus 5’s pass rate is roughly 1.5× that of the next-best model for the same cost per task, and even at its lowest effort setting it passes more tasks than any other model.
- OSWorld 2.0: Opus 5 outperforms every other model at a given cost, surpassing Fable 5’s best result at a bit over a third of the cost.
Opus 5 is also the most cost-efficient model on several additional evaluations, including ARC-AGI 3, GDPval-AA v2, OSWorld 2.0, HLE, AutomationBench, and DeepSearchQA.
Scientific and life-sciences improvements
Compared with Opus 4.8, Opus 5 provides meaningful improvements on the company’s life-sciences evaluations covering structural biology, organic chemistry, and bioinformatics. Notable gains include:
- Inferring molecular structures from spectroscopy: an internal benchmark shows Opus 5 scoring 10.2 percentage points higher than Opus 4.8.
- Predicting how protein sequence variations affect function: Opus 5 scores 7.7 percentage points higher.
Stronger visual outputs
Anthropic highlights improved visual capabilities in Opus 5, with examples such as wind-tunnel visualizations of air flow and simplified, interactive cell illustrations.
How Opus 5 works in practice
Opus 5 is described as better at verifying its own work and iterating until success. Early-access testing and customer reports include concrete examples:
- In one Frontier-Bench task the model reconstructed a machine part as a 3D FreeCAD model despite having no direct way to view the drawing; it wrote its own computer-vision pipeline to extract geometry from raw pixels and succeeded repeatedly while competing models failed under the same setup.
- Faced with a real bug in a popular open-source package manager, Opus 5 identified the root cause and fixed an edge case that the community patch missed; a competing model only addressed the surface symptom.
- An engineer used Opus 5 to build a market data feed for a new exchange in a single session; lacking a live feed for validation, Opus 5 built its own test harness to verify parsing correctness.
Multiple corporate partners reported improvements across workflows such as code review, financial research, legal agent tasks, presentation authoring, genomics analysis, and frontend development. Testimonials from organizations including Cursor, Zapier, Box, JetBrains and others emphasize higher accuracy, reduced resource use, and greater consistency compared with prior Opus models.
Alignment and safety
Anthropic reports that in pre-deployment automated behavioral audits Opus 5 is the company’s most aligned model to date, exhibiting the lowest rates of deceptive behavior and the least susceptibility to being tricked into misuse. On the audit, Opus 5 scores 2.3 for overall misaligned behavior, the lowest among recent models.
The company also states that Opus 5 does not advance the frontier in risky, dual-use capabilities. Evaluations conducted with private-sector and government partners indicate that Mythos 5 remains stronger for both biology research and offensive cybersecurity. While Opus 5 and Mythos 5 perform similarly at identifying software vulnerabilities, Opus 5 is substantially less successful at developing exploits.
This divergence is illustrated by OSS-Fuzz evaluations: Opus 5 approaches Mythos 5 on finding vulnerabilities but scores far lower on exploit development.
Safeguards and operational controls
Opus 5’s safeguards resemble those used for Opus 4.8, with some stronger guardrails for specific cyber tasks. Cybersecurity classifiers for Opus 5 are proportionally less restrictive than those for Fable 5: they permit source-code vulnerability discovery but block binary-based scanning, penetration testing, and exploit generation. Anthropic expects classifier interventions to occur about 85% less often than for Fable 5.
On Claude.ai, Claude Code, and Claude Cowork, flagged requests default to falling back to Opus 4.8. Fallbacks to Opus 4.8 can also be enabled via the API. The Cyber Verification Program (CVP) provides enterprises and approved researchers with a version of Opus 5 that has fewer security restrictions for legitimate cybersecurity work.
For biology-related uses, Opus 5 carries a similar set of safeguards as Opus 4.8. As a result, Opus 5 becomes Anthropic’s most capable generally available model for scientific research, though it still has limitations on long-running, autonomous research tasks where the company expects the greatest biology-related risks; Mythos 5 remains stronger for such work.
Availability and pricing
Claude Opus 5 is available immediately across all platforms and is priced the same as Opus 4.8: $5 per million input tokens and $25 per million output tokens. Developers can call claude-opus-5 via the Claude API.
Opus 5 is also offered in a Fast mode that runs roughly 2.5× the default speed; Fast mode costs twice Opus 5’s base price on the Claude Platform and is available via usage credits in Claude Code.
Two beta features are released alongside Opus 5:
- Mid-conversation tool changes on the Claude Platform, which let developers change what tools Claude can use within a conversation without invalidating the prompt cache.
- Automatic fallbacks on the API, which route safety-flagged requests on Opus 5 (or Fable 5) to another model automatically so requests are routed to the best available model rather than being blocked.
Anthropic notes that, consistent with prior Opus models, Opus 5 does not impose data-retention requirements for general access, and the company provides a prompting guide to help developers get the best results.
Footnote on benchmarks
Some reported results (for example Frontier-Bench v0.1) come from internal runs: mini-SWE-agent harness, a GKE backend, and mean reward over five attempts per task. Opus 4.8 served as the fallback on safety-classifier refusals for Opus 5 and Fable 5.
(Detailed technical and safety documentation including a System Card accompanies the release.)
