Anthropic launches Claude Opus 5: advanced intelligence at half the cost

Anthropic has announced the release of Claude Opus 5, a new artificial intelligence model that is already available to users. The new model is described as thoughtful and proactive, approaching the flagship Claude Fable 5 in terms of intelligence, but at half the cost.

Key facts about the release:

— Cost and performance: Opus 5 sets a new standard in programming and intellectual labor tasks on benchmarks such as Frontier-Bench and GDPval-AA. At the same time, the cost remains at the level of the previous Opus 4.8 version: $5 per one million input tokens and $25 per one million output tokens.

— Programming and science: On the Frontier-Bench v0.1 benchmark, the new model outperforms Opus 4.8 by more than two times at a lower cost per task. The model also demonstrates significant improvements in the field of natural sciences: for example, in organic chemistry, the accuracy of deriving molecular structures from spectroscopic data increased by 10.2 percentage points.

— Autonomy and creativity: Opus 5 has become better at verifying its own results and finding non-standard solutions. In one test on building a 3D model, without a direct way to "see" the provided drawing, the AI independently wrote a computer vision pipeline to extract geometry from raw pixels and successfully recreated the part. The model is also capable of creating complex visual elements, such as interactive cell illustrations and wind tunnel simulations.

— Safety and constitution: Automated auditing showed that Opus 5 is Anthropic's most "consistent" model to date. It adheres better to the Claude Constitution than Opus 4.8, Sonnet 5, or Fable 5, and demonstrates the lowest propensity for deceptive behavior.

— Cybersecurity: In the field of cybersecurity, the model still lags behind Mythos 5, especially in exploit development. However, Opus 5's protective classifiers are less strict than those of Fable 5 and will block requests approximately 85% less often.

Claude Opus 5 has become the default model in the Claude Max corporate subscription and the most powerful option for Claude Pro users. The model also supports Fast mode, which operates approximately 2.5 times faster than the base speed but costs twice as much.

Along with the release, Anthropic added two beta features for the API: the ability to change available AI tools during a conversation without resetting the cache, and an automatic redirection system that routes requests blocked by security filters to another model.

Source: anthropic.com