Anthropic has released its new flagship model, Claude Opus 5. It has been available on all of the company’s platforms since July 24, 2026: For Claude Max subscribers, Opus 5 is the new standard model; in Claude Pro, it can be selected as the most powerful option; and developers can access it via the Claude API. According to Anthropic, the model comes close to the intelligence level of Claude Fable 5—but costs only about half as much.
Key Points at a Glance
- Claude Opus 5 has been available on all Anthropic platforms and via the API since July 24, 2026.
- Standard model in Claude Max, top option in Claude Pro.
- Nearly on par with Claude Fable 5—at about half the cost per task.
- A new “Effort” setting controls how much computational effort the model invests.
- Pricing remains unchanged: $5 per million input tokens, $25 per million output tokens.
Nearly Fable 5 performance at half the price
Anthropic positions Claude Opus 5 as a model for demanding programming, analysis, and knowledge work. The real highlight lies less in a single benchmark record than in the performance-to-cost ratio: Opus 5 is designed to handle the same workload as its predecessor, Opus 4.8, but costs less per task and, in many areas, comes close to matching the more expensive Fable 5.
The new “Effort” setting helps with this. Users can use it to determine for themselves how much computational power the model devotes to a query. Those who need fast, affordable answers can turn the slider down; for tricky tasks, the full computational power can be tapped into. This way, everyone can determine their own balance between speed, cost, and quality.
Anthropic recently cleared up the access chaos surrounding its larger sister model. If you want to learn more about its place in the portfolio, you can find the details in our article on Claude Fable 5.
Benchmarks: A Clear Lead at Lower Costs
For the launch, Anthropic is presenting a whole range of metrics. On Frontier-Bench v0.1, a test for software engineering, Opus 5 takes the top spot and doubles the score of Opus 4.8—at a lower cost per task. On CursorBench 3.2, the model trails Fable 5’s best score by only 0.5 percent at maximum computational effort, yet it also requires only about half the cost.
Anthropic also aims to lead beyond programming. On ARC-AGI 3, a test for novel problem-solving scenarios, Opus 5 achieves three times the score of the second-place model, according to the manufacturer. In the OSWorld 2.0 computer-use benchmark, it outperforms the competition at every cost point and surpasses Fable’s best score at about one-third of the cost. It also achieves top scores on Zapier’s AutomationBench, GDPval-AA, and other evaluations.
Progress is also evident in specialized fields: In internal life sciences tests, Opus 5 excels particularly in organic chemistry (10.2 percentage points above Opus 4.8) and in tasks related to proteins (7.7 percentage points higher). Early test customers also report improvements in debugging, data analysis, financial research, and long-term programming projects with many sequential steps.
Security: The Most Security-Focused Claude Model to Date
Anthropic describes Opus 5 as its model most strongly optimized for safety and policy compliance to date. In an automated behavioral audit, it achieved a score of 2.3—the lowest rate of misconduct among current models—and, according to the company, adheres more strictly to Claude’s constitution than Opus 4.8, Sonnet 5, or Fable 5.
The company remains cautious on the issue of dual-use applications. Opus 5 deliberately does not set any new benchmarks in this area and continues to lag behind the Mythos 5 model, particularly in biological research and offensive cybersecurity. Although Opus 5 now identifies security vulnerabilities almost as well as Mythos 5, it is significantly weaker at turning these vulnerabilities into functional attacks. Anthropic has therefore opted not to provide targeted training for offensive cybersecurity tasks.
You are currently viewing a placeholder content from YouTube. To access the actual content, click the button below. Please note that doing so will share data with third-party providers.
Specifically, new classifiers still allow the analysis of source code to detect vulnerabilities but block the scanning of binary files, penetration testing, and the generation of exploits. Such classifiers are designed to intervene approximately 85 percent less frequently than with Fable 5. If a request is blocked, Claude AI, Claude Code, and Claude Cowork automatically downgrade to Opus 4.8 by default.
Two New Beta Features for Developers
There are two new features in beta for developers. First, tools can be swapped within an ongoing conversation without losing the prompt cache. Second, requests blocked by the safety filters can be automatically forwarded to another model. This means that API requests are no longer completely rejected but, in case of doubt, are routed to the next available model.
Pricing and Availability
The pricing structure remains unchanged. The API continues to cost $5 per million input tokens and $25 per million output tokens—exactly the same as in Opus 4.8. In addition, there is an optional Fast Mode that performs calculations about 2.5 times faster but costs twice as much.
| Feature | Claude Opus 5 |
|---|---|
| Available since | July 24, 2026 |
| Platforms | Claude Max (Standard), Claude Pro (most powerful option), Claude API |
| API Price Input | $5 / 1 million tokens |
| API Price (Output) | $25 / 1 million tokens |
| Fast Mode | Approx. 2.5× faster, double the cost |
| Special Feature | Effort settings, automatic fallbacks, mid-conversation tool switching |
Conclusion
With Claude Opus 5, Anthropic delivers not so much a spectacular leap to the absolute top as a very practical model for everyday use: nearly Fable 5-level performance, but at half the price and with an effort setting that puts control over costs and speed in the users’ hands. The fact that pricing and security architecture remain largely stable makes the transition straightforward for existing users. It remains to be seen just how significant the advantage will be in everyday use—but the direction is definitely the right one. For a broader context regarding Anthropic’s infrastructure, it’s also worth taking a look at the 2-gigawatt deal between AMD and Anthropic.
Sources: Anthropic
