{"id":70603,"date":"2026-07-25T00:44:16","date_gmt":"2026-07-24T22:44:16","guid":{"rendered":"https:\/\/www.apfelpatient.de\/?p=70603"},"modified":"2026-07-25T01:06:39","modified_gmt":"2026-07-24T23:06:39","slug":"claude-opus-5-anthropic-new-ai-model","status":"publish","type":"post","link":"https:\/\/www.apfelpatient.de\/en\/news\/claude-opus-5-anthropic-neues-ki-modell","title":{"rendered":"Claude Opus 5: Anthropic's new AI model is here"},"content":{"rendered":"

Two months after its predecessor, Anthropic is releasing the next generation of Opus \u2013 at the same price. The model is intended to closely match the company's flagship model and will become the standard setting for paying customers. At the same time, the company is deliberately limiting its capabilities in one specific area.<\/strong><\/p>\n\n\n\n

Anthropic released Claude Opus 5 on Friday, replacing Opus 4.8 from May<\/a>. The roughly two-month gap reflects the company's recent release schedule. Opus represents the middle tier of several models: below it are the smaller Haiku and Sonnet, and above it is the Mythos class with Fable 5. According to the company, the new model is intended to approach the performance of Fable 5 while costing half as much.<\/p>\n\n\n\n

What the manufacturer presents in terms of measured values<\/h2>\n\n\n\n

The published figures are all from Anthropic's own tests and have not been independently verified. According to these figures, Opus 5 doubles the performance of its predecessor on a software development benchmark while reducing the cost per task. On a test for novel problems, the result is said to be three times higher than that of the next best model.<\/p>\n\n\n\n

One striking detail regarding its operation is that Anthropic describes several instances where the model itself reconstructed missing tools \u2013 for example, image recognition to analyze a technical drawing to which it had no direct access. In the life sciences, the company cites specific gains compared to its predecessor, including a 10.2 percentage point improvement in the analysis of spectroscopy data.<\/p>\n\n\n\n

More reliable than such manufacturer specifications are feedback from development companies that had early access. This feedback is also positive, but it comes from partners that Anthropic selected for the announcement.<\/p>\n\n\n\n

Price, speed, and who gets it<\/h2>\n\n\n\n

The billing remains unchanged at $5 per million incoming and $25 per million outgoing tokens \u2013 identical to the previous model. An optional faster mode runs at approximately two and a half times the speed and costs twice as much.<\/p>\n\n\n\n

For end users, the default setting is changing: Opus 5 will become the standard model in the most expensive subscription and simultaneously the most powerful model in the mid-range plan. Two new features for developers will be released simultaneously as beta versions, including the ability to switch tools mid-conversation without losing the cache.<\/p>\n\n\n\n

Where the model appears on Apple devices<\/h2>\n\n\n\n

For readers of this page, the most relevant question is where the model ends up. Anthropic's developer tool, Claude Code, runs as a Mac app and now tests iOS apps directly in the simulator<\/a>. The office application Claude Cowork has been available on iPhone and iPad since the beginning of July<\/a>. Both use the current generation of the model.<\/p>\n\n\n\n

However, there is no connection to Apple's own assistant. Siri AI is based on a modified version of Google's Gemini, not on Claude.<\/p>\n\n\n\n

Why Opus 5 deliberately stays behind the top model<\/h2>\n\n\n\n

The most revealing part of the announcement concerns its limitations. Anthropic explicitly states that Opus 5 offers no advancements in risky, dual-use capabilities and falls short of Mythos 5 in both biological research and offensive cybersecurity. According to the company, the model was intentionally not trained for cyber tasks.<\/p>\n\n\n\n

While it comes close to the top-of-the-line model in detecting software vulnerabilities, it lags significantly behind in developing attack code to exploit these vulnerabilities. This is precisely where additional filters come into play: they allow searching for vulnerabilities in the source code but block penetration tests and the generation of attack code. If a query is flagged, the system reverts to its predecessor, Opus 4.8.<\/p>\n\n\n\n

The backstory that explains this reluctance<\/h2>\n\n\n\n

This reluctance has a specific reason. On June 12 \u2013 three days after the release \u2013 the US Department of Commerce instructed Anthropic to block Fable 5 and Mythos 5 for foreign nationals<\/a>. This was triggered by reports of circumvention of cybersecurity protection mechanisms. Anthropic subsequently disabled both models for all customers.<\/p>\n\n\n\n

The order affected users in Germany, Austria, and Switzerland equally \u2013 unlike regulations from Brussels, EU membership played no role here; the sole determining factor was citizenship outside the US. Nineteen days later, the Department lifted the restrictions, and both models returned on July 1st<\/a>.<\/p>\n\n\n\n

Against this backdrop, Opus 5 reads as a model that deliberately remains below the threshold that led to the shutdown in June. Anthropic doesn't say whether this interpretation is correct \u2013 however, the description of its own limitations in the announcement is unusually detailed.<\/p>\n\n\n\n

The gap between the model levels is shrinking<\/h2>\n\n\n\n

What's remarkable about this release<\/a> isn't so much the performance leap as the pricing. A model that, according to the manufacturer, comes close to its own top-of-the-line counterpart, costs exactly the same as its predecessor. The differentiation that manufacturers have built up over the years is thus losing its precision \u2013 at least as long as the measurements come from the companies evaluating their own models. (Image: Shutterstock \/ Stockinq)<\/p>\n\n\n\n