Anthropic has launched Claude Opus 5, now the default option on Claude Max.
Opus 5 is priced the same as its predecessor, Opus 4.8: USD $5 per million input tokens and USD $25 per million output tokens. Anthropic said it is the strongest option on Claude Pro and is priced below Claude Fable 5.
Anthropic presented the release as an advance in coding and knowledge work, citing benchmark results that place Opus 5 ahead of rivals on Frontier-Bench and GDPval-AA. The model remains behind Mythos 5 on cybersecurity tasks, and Anthropic said it was not trained specifically for cyber work.
Cost is central to the launch. Anthropic said Opus 5 comes close to Claude Fable 5's performance at about half the cost per task in several tests, while matching Opus 4.8 on price. The model also includes an effort setting that lets users trade off reasoning depth against token use, speed and cost.
On software engineering tests, Anthropic said Opus 5 more than doubled Opus 4.8's performance on Frontier-Bench v0.1 while reducing cost per task. On CursorBench 3.2, the model came within 0.5% of Fable 5's peak score at maximum effort, again at half the cost per task.
In knowledge work and problem-solving, Anthropic pointed to ARC-AGI 3, where Opus 5 scored three times higher than the next-best model, according to the company. It also said the model led other systems on Zapier AutomationBench and OSWorld 2.0 at comparable or lower cost levels, including at the lowest effort setting.
Research focus
Anthropic also positioned Opus 5 as a stronger tool for scientific research, saying it improved on every life sciences evaluation the company uses internally, including structural biology, organic chemistry and bioinformatics.
On internal organic chemistry testing based on inferring molecular structures from spectroscopy data, Opus 5 scored 10.2 percentage points higher than Opus 4.8, Anthropic said. It added that the model was 7.7 percentage points higher on protein sequence variation predictions.
The launch materials also emphasised the model's behaviour on extended tasks. Anthropic said Opus 5 is better at checking its own work and revising output until a task is complete, citing examples including building a computer vision pipeline, identifying and fixing a bug in an open-source package manager, and constructing a market data feed for a new exchange in a single session.
Several external users published assessments alongside the launch. Their comments focused on debugging, legal work, finance, software development and scientific analysis, with many comparing Opus 5 directly with Opus 4.8 and, in some cases, with Fable 5.
Scott Wu, Chief Executive Officer, said the model performs strongly on coding work. "On FrontierCode 1.1, Claude Opus 5 approaches Fable-level performance at half the cost. Within Devin, it shows particular strength on difficult debugging and root-cause analysis tasks," Wu said.
Sualeh Asif, Co-Founder, drew a similar comparison on another benchmark. "Claude Opus 5 delivers near Fable 5 intelligence at Opus speed and cost. On CursorBench it's just under Fable 5 and has many of the same behaviors," Asif said.
Wade Foster, Chief Executive Officer, highlighted workflow automation. "Claude Opus 5 topped Zapier's AutomationBench leaderboard without spending more tokens than prior Claude models, hitting 100% on full churn-prevention sequences," Foster said.
In the life sciences, Alfredo Andere, Chief Executive Officer, said the model showed more caution in analysis. "On genomics analysis work, Claude Opus 5 behaves more like a careful scientist, reaching for correct statistical tests to rule out confounders and cross-checking results," Andere said.
Safety measures
Anthropic said behavioural audits show Opus 5 is its most aligned model so far, with an overall misaligned behaviour score of 2.3. The company said the model follows its constitution more closely than Opus 4.8, Sonnet 5 or Fable 5, and shows lower rates of deceptive behaviour and misuse susceptibility.
At the same time, Anthropic said Opus 5 does not advance the frontier in dual-use risk areas. It remains behind Mythos 5 in biology research and offensive cybersecurity, and substantially behind it in exploit development on OSS-Fuzz benchmarks, according to the company.
To limit misuse, flagged requests in Claude.ai, Claude Code and Claude Cowork will automatically fall back to Opus 4.8 by default, Anthropic said. The company added that cybersecurity controls block binary-based scanning, penetration testing and exploit generation, while approved researchers in its Cyber Verification Program can access versions with lower restrictions.
Anthropic also said biology-related requests blocked on Fable 5 will automatically route to Opus 5, which it described as its primary generally available model for scientific research. Fast Mode is also being offered at 2.5 times the default speed for double the base price.
Deepak Singh, Vice President of Agentic AI, said Opus 5 "deeply understands codebases and holds threads across multi-step tasks in Kiro."