Claude Opus 4.8: What's new about the AI model Apfelpatient

Just six weeks after the last update, Anthropic is releasing the next version of its AI model, Claude Opus 4.8 – at the same price, but with noticeably improved judgment and a striking focus on honesty. The launch is accompanied by three new features, primarily aimed at demanding, long-term tasks.

Anthropic is picking up the pace: With Claude Opus 4.8, the next iteration is released just six weeks after its predecessor, Opus 4.7. For the Apple world, this is more than just a footnote, as Claude is among the AI models that Apple is reportedly testing for a future Siri generation – meaning that, in the long run, users could one day choose between Claude, Gemini, and other providers as the AI backend for Siri. This makes it all the more revealing to examine what the new version actually improves – and where Anthropic has focused its efforts on reliability.

What makes Opus 4.8 better

Anthropic describes Opus 4.8 as a "more effective employee" with sharper judgment, greater honesty about their own work progress, and the ability to work independently for longer than its predecessors. The price remains the same as Opus 4.7.

The comparative figures published by the company show a consistent, albeit moderate, increase. For agentic programming, the score climbs from 64.3 to 69.2 percent, and for cross-disciplinary reasoning with tools, from 54.7 to 57.9 percent. Agentic computer use improves from 82.8 to 83.4 percent, and agentic financial analysis from 51.5 to 53.9 percent. The largest jump is recorded in the score for knowledge-based work, which, according to Anthropic, rises from 1,753 to 1,890 points. In early adopter tests, the model is also being compared against competing models such as GPT-5.5.

Honesty as a key innovation

The most noticeable improvement concerns not so much pure performance metrics as the model's behavior. Anthropic particularly emphasizes the honesty of Opus 4.8. A well-known problem with AI models is that they sometimes jump to conclusions and confidently claim progress, even when the evidence is weak.

This is precisely where the new version is intended to be more cautious. Early testers report that Opus 4.8 more frequently points out uncertainties in its own work and makes unsubstantiated claims less often. In their own analyses, the model is about four times less likely than its predecessor to let errors in user-written code pass without comment.

Anthropic also draws a positive conclusion regarding the safety assessment. The internal alignment team certifies that Opus 4.8 has achieved new record highs in prosocial attributes such as supporting user autonomy and acting in the users' best interests. Rates of misguided behavior are significantly lower than those of Opus 4.7 and thus on par with the company's best-aligned model to date.

Faster and cheaper in Fast Mode

In addition to improving content, Anthropic is also focusing on speed and cost. The so-called Fast Mode, in which the model operates at increased speed, is now approximately 2.5 times faster than before. At the same time, according to the company, this accelerated mode costs three times less than previous models.

The regular pricing structure remains unchanged: For normal use, Anthropic continues to charge $5 per million incoming tokens and $25 per million outgoing tokens. In Fast Mode, the prices are $10 and $50, respectively. Developers access the model via the API under the name claude-opus-4-8.

Three new features at launch

Alongside the model itself, Anthropic is launching three additional features. The first is dynamic workflows, initially available as a research preview. These allow Claude, within the Claude Code development tool, to tackle very large tasks: The model plans the work, then starts hundreds of parallel sub-agents in a single session and checks the results before reporting back. Anthropic cites migrations of hundreds of thousands of lines of code, from initiation to merge, as an example. This feature is available in the Enterprise, Team, and Max plans.

The second new feature is effort control in Claude.ai and Cowork. Directly next to the model selection, you can now specify how much effort Claude puts into a response. At higher levels, the model thinks more frequently and thoroughly; at lower levels, it responds faster and uses up resources more slowly. By default, Opus 4.8 operates at a high effort level; even more intensive levels are available for particularly difficult tasks. This control is available in all plans.

The third change is aimed at developers: The Messages interface now accepts system entries within the message array. This allows instructions to be updated to Claude mid-task without destroying the cached context or having to inject the change via a user action.

What's planned next

Anthropic describes Opus 4.8 as a noticeable, but moderate, improvement over its predecessor – and also hints at what the company is already working on. For one thing, future models are planned that will offer many of Opus's capabilities at a lower cost.

Secondly, a new model class is planned that will offer an even higher level of performance than Opus. As part of an initiative called Project Glasswing, a small number of organizations are already using a preview model called Claude Mythos Preview for cybersecurity tasks. According to the company, models of this performance level require stronger safeguards before they can be made generally available – these are expected to follow in the coming weeks.

Anthropic's pace is increasing

What's remarkable about this release isn't so much a single metric as the pace behind it. With only six weeks between two Opus versions, Anthropic significantly shortens its update cycle, delivering not only higher benchmark scores but also a model that is more transparent about its own limitations.

For Apple users, the most intriguing question remains: what role will such models play in the company's ecosystem in the future? Should Apple implement its plans and integrate external AI providers more deeply into Siri, the quality of models like Opus 4.8 will ultimately determine how convincing such an assistant is. (Image: Anthropic)