Key Takeaways
- Anthropic has launched Claude Opus 5 across all platforms, retaining the same API pricing as Opus 4.8 ($5 / M input tokens, $25 / M output tokens) while claiming notable performance gains in coding, knowledge work, computer use, and scientific research.
- Opus 5 is positioned as a cost‑effective everyday alternative to the more capable Claude Fable 5, approaching Fable 5’s performance at roughly half the price, though it still trails Fable 5 and the specialized Mythos 5 on certain benchmarks.
- The model introduces an effort‑setting knob that lets users trade additional reasoning and token usage for speed and cost, enabling organizations to reserve higher effort for complex tasks.
- Benchmark results show Opus 5 outperforming Opus 4.8 and often beating competing models on agentic work (Frontier‑Bench), knowledge‑work tasks (GDPval‑AA v2), and business‑workflow completion (AutomationBench), while lagging slightly on some deep‑reasoning and security‑focused tests.
- Safety and cyber‑security controls are strengthened: Opus 5 improves vulnerability detection but remains limited in exploit generation, with classifiers blocking binary‑based scanning and exploit creation far more often than those on Fable 5.
- Additional beta features include automatic fallback to Opus 4.8 for flagged requests, a tool‑swap capability that preserves prompt caches, and a Fast Mode that runs Opus 5 at ~2.5× standard speed for twice the base price.
- General access to Opus 5 carries no data‑retention requirements, and existing members of the Cyber Verification Program receive a less‑restricted version for approved security work.
Overview of the Opus 5 Release
Anthropic has made Claude Opus 5 available on all its platforms, keeping the base API pricing unchanged from its predecessor Opus 4.8 at $5 per million input tokens and $25 per million output tokens. Developers can call the model via the Claude API using the identifier claude-opus-5. Opus 5 now serves as the default model on Claude Max and is the strongest offering available through Claude Pro. The release is framed as an everyday alternative to the more powerful Claude Fable 5, with Anthropic claiming that Opus 5 approaches Fable 5’s performance while costing roughly half as much.
Pricing and Positioning Relative to Fable 5
Although Opus 5 shares the same token‑level pricing as Opus 4.8, Anthropic emphasizes its cost‑efficiency advantage over Fable 5. The company states that Opus 5 can achieve performance comparable to Fable 5 on many tasks at about ½ the price, making it a viable “workhorse” model for regular use. Nevertheless, Anthropic acknowledges that Opus 5 still lags behind Fable 5 on certain benchmarks and remains inferior to the specialized Mythos 5 model in areas such as offensive cybersecurity and long‑running biological research.
Effort‑Setting Flexibility
A notable feature of Opus 5 is the adjustable effort setting, which lets users trade additional reasoning steps and token consumption for faster, cheaper responses. By dialing effort up or down, organizations can allocate higher computational budgets only to complex, high‑value requests while using lower‑effort, cheaper modes for routine queries. This flexibility aims to prevent overpaying for maximum capability on every interaction.
Performance Gains in Coding and Agentic Work
Anthropic’s strongest performance claims center on agentic work, where the model must execute a sequence of actions rather than output a single answer. On the Frontier‑Bench v0.1 suite (using the mini‑SWE‑agent harness on a Google Kubernetes Engine backend), Opus 5 scored 43.3%, outperforming Fable 5 (33.7%) and Opus 4.8 (21.1%). The benchmark also placed GPT‑5.6 Sol at 34.4%. The company notes that Opus 4.8 served as a fallback when safety classifiers refused a request, meaning the headline score reflects some assistance from the older model.
On CursorBench 3.2, Opus 5 running at its maximum effort setting came within 0.5% of Fable 5’s peak result while costing half as much per task. Anthropic further claims that Opus 5 delivers superior performance at a given cost across its high, xhigh, and maximum effort settings compared with competing models.
Knowledge‑Work and Computer‑Use Evaluations
Additional internal evaluations show Opus 5 advancing knowledge‑work and computer‑use capabilities. On GDPval‑AA v2, Opus 5 achieved a score of 1,861, surpassing Fable 5 (1,747), Opus 4.8 (1,593), and GPT‑5.6 Sol (1,736). In the AutomationBench, which measures end‑to‑end completion of business workflows, Opus 5 posted a 26% pass rate—substantially higher than Fable 5 (17.4%), Opus 4.8 (17%), and GPT‑5.6 Sol (18.1%). Anthropic states that this pass rate is roughly 1.5 times that of the next‑best model at equivalent cost per task, and that Opus 5 exceeds all competitors even at its lowest effort setting.
Limitations on Deep Reasoning and Security Benchmarks
Advantages are not uniform across all tests. On DeepSWE v1.1, Opus 5 scored 68.8%, trailing Fable 5 (69.7%) and GPT‑5.6 Sol (72.7%). In the Humanity’s Last Exam (without tools), Opus 5 earned 56.3%, just shy of Fable 5’s 56.5%. These results indicate that while Opus 5 excels at iterative, agentic tasks, it remains slightly behind the top models on certain deep‑reasoning assessments.
Illustrative Use‑Case Examples
Anthropic supplies qualitative examples to illustrate Opus 5’s strengths. One case describes the model reconstructing a machine part as a 3D FreeCAD model after building a computer‑vision pipeline to extract geometry from raw pixel data. Another example shows Opus 5 diagnosing and fixing the root cause of a bug in an open‑source package manager, whereas a prior model only addressed the immediate symptom. An early‑access engineer used Opus 5 to construct a market‑data feed prototype for a new exchange, creating a test harness when live data was unavailable. These anecdotes are presented as stemming from evaluations and early‑access deployments, separate from the formal benchmark numbers.
Scientific Research Improvements
In life‑science evaluations covering structural biology, organic chemistry, and bioinformatics, Opus 5 outperformed Opus 4.8. The largest gain was a 10.2‑percentage‑point improvement on an internal task that required inferring molecular structures from spectroscopy data. Opus 5 also scored 7.7 percentage points higher on tasks predicting how protein‑sequence alterations affect function. Anthropic positions Opus 5 as its most capable generally available model for scientific research, while noting that it remains unsuitable for fully autonomous, long‑running biological work—where the specialized Mythos 5 still leads.
Cyber‑Security Controls and Dual‑Use Safeguards
Regarding risky dual‑use capabilities, Anthropic asserts that Opus 5 does not advance the frontier in harmful areas. In private‑ and public‑sector evaluations, Opus 5 lagged behind Mythos 5 in biology research and offensive cybersecurity. On the OSS‑Fuzz test (identifying and exploiting vulnerabilities without extensive human guidance), Opus 5 passed 79.4% of vulnerability‑identification challenges—close to Mythos 5’s 80% and ahead of Opus 4.8’s 61.5%. However, its ability to turn those findings into working exploits was markedly lower: Mythos 5 solved 13 exploitation challenges, Opus 5 managed only 4, and Opus 4.8 solved none.
Anthropic says it did not specifically train Opus 5 on cyber tasks; improvements arose incidentally from broader capability gains, prompting tighter controls for a narrow set of security‑related requests. Opus 5 can be used to locate source‑code vulnerabilities, but its classifiers block binary‑based scanning, penetration testing, and exploit generation. These safeguards are expected to trigger ≈85 % less often than those attached to Fable 5.
When a request is flagged by Opus 5 or Fable 5 in Claude.ai, Claude Code, or Claude Cowork, the system falls back to Opus 4.8 by default. API customers can enable an automatic fallback beta feature that routes flagged requests to another available model instead of outright blocking them. Members of Anthropic’s Cyber Verification Program receive immediate access to a less‑restricted Opus 5 variant for approved security work.
Additional Beta Updates and Fast Mode
Two beta updates accompany the Opus 5 release. The first is the automatic fallback mechanism described above. The second allows developers to swap the tools available to Claude during a conversation without invalidating the prompt cache, facilitating dynamic workflow adjustments.
Anthropic also introduces a Fast Mode that runs Opus 5 at roughly 2.5 × its standard speed. Fast Mode costs twice the base price on the Claude Platform and can be accessed via usage credits in Claude Code. General access to Opus 5 carries no data‑retention requirements, simplifying compliance for users who do not opt into premium features.
Behavioral Safety and Misaligned‑Behavior Scores
A separate behavioral audit awarded Opus 5 an overall misaligned‑behavior score of 2.3, the lowest among Anthropic’s recent models. The audit indicates reduced rates of deceptive behavior and fewer reckless actions with potentially irreversible consequences compared to Opus 4.8, Sonnet 5, and Fable 5. This improvement reinforces Anthropic’s claim that Opus 5 balances heightened capability with stronger safety safeguards.
