Anthropic releases Claude Opus 5! Performance close to Fable 5’s, but priced at half—becomes the latest default model

Well-known AI startup Anthropic officially released its new model Claude Opus 5 on July 24 (Taipei time) today. The model is centered on “top-tier intelligence, affordable pricing.” Not only has it set new records in programming and research evaluations, its cost is also only half that of the leading model Fable 5. Starting today, Opus 5 will become the new default model for Claude Max, bringing the public a more efficient everyday AI experience.
（Background: Anthropic no longer shows Claude’s summarized thought process to prevent Fable 5 from being distilled）
（Background add-on: Anthropic is pushing for an IPO! Reports say it would require employees to sign a “10b5-1 automatic stock selling” plan to prevent a selloff wave at a 10B valuation）

Table of Contents

Toggle

• Breakthrough in programming and agent capabilities, achieving maximum efficiency at low cost
• Most aligned, with flexible relaxation of network security vulnerability-finding restrictions
• New Fast mode launched, API support for automatic downgrade

Amid the ever-accelerating pace of AI model iterations, Anthropic throws another shockwave. On July 24 (Taipei time), Anthropic officially announced the launch of its new model Claude Opus 5. Positioned as a “deliberate and proactive” next-generation model, its overall intelligence performance has already come close to the company’s top-tier frontier model Claude Fable 5, but its usage price is cut in half—demonstrating extremely high cost efficiency.

Breakthrough in programming and agent capabilities, achieving maximum efficiency at low cost

Opus 5 reached new industry-standard (state-of-the-art) benchmarks across multiple test suites. For software engineering tasks, it not only surpasses all competitors in the Frontier-Bench tests, but also significantly reduces the cost per task—boosting the performance of the previous Opus 4.8 by more than double. In knowledge work tasks such as the commercial tasks in Zapier AutomationBench, the pass rate at the same cost is 1.5 times that of the second-best model.

In addition, Opus 5 demonstrates strong “agentic capabilities,” enabling it to autonomously iterate and verify work. For example, without directly viewing blueprints, it can build a complete 3D geometric model from code; and within a single conversation, it can independently debug a real system and establish a validation framework. These are achievements that competing models had difficulty reaching in early testing.

Most aligned, with flexible relaxation of network security vulnerability-finding restrictions

In terms of safety and alignment, Opus 5 sets the best record among recent models. According to automated behavior audits, its overall misalignment behavior score is only 2.3. Not only is it more compliant with the constitution than the previous Opus 4.8 and even Fable 5, but its deception rate has also been reduced to the lowest.

For the highly watched cybersecurity domain, Anthropic chose to slightly loosen the classifier restrictions, allowing Opus 5 to help “find vulnerabilities” in the source code. The expected frequency of classifier intervention is about 85% lower than with Fable 5. However, to prevent malicious exploitation, the official still strictly blocks the generation of vulnerability scanning or exploit code.

New Fast mode launched, API support for automatic downgrade

To enable more flexible deployment for enterprises and developers, Opus 5 has become the new default model for Claude Max starting today, and it is also the most powerful option in the Claude Pro plan. The official also simultaneously introduced “Fast mode,” which runs at 2.5 times the default speed, with pricing set at twice the base price.

On the developer experience side, Claude Platform added support for changing tools mid-conversation without invalidating the prompt cache. At the same time, the API also introduced an “auto-fallback” mechanism: when a request is flagged by the safety classifier, the system automatically routes it to another model (such as Opus 4.8) to ensure that service remains smooth and uninterrupted.