TechCrunch reports that Anthropic's new language model, Opus 5.5, was released on Tuesday and has set a new record for performance in the programming and knowledge domain. The company says the new model outperforms the larger Fable model in many ways and succeeds in a number of informal tasks that the Fable couldn't handle. In addition, the new model is significantly cheaper than its predecessor. The token price for Opus 5.5 will be $20 per million tokens, while the previous model cost $25. Other metrics also showed similar price declines. The model runs faster, which reflects the overall reduction in computing resources required to maintain it. The new version also makes significant changes to the way Opus communicates: Opus 5.5 now uses jargon less often and places important information at the beginning of its messages more often. The launch of the model took place just two months after the release of Opus 5, which was presented on July 24. According to the announcement, Sonnet 5.5 and Haiku 5.5 will be released "in the coming weeks" with similar performance improvements. Anthropic says Opus 5.5 is comparable to Myth
os in its biology and cybersecurity capabilities, so its release is subject to the same security measures as the Fable model. These measures limit the use of models to detect vulnerabilities in compiled programs or develop recognizable biological weapons, among other tasks. Opus 5.5 is the first model released by Anthropic since CEO Dario Amodei made decisions to slow down the development of AI capabilities to match the pace of progress in alignment. “I have concluded that fully addressing risk requires even greater caution,” Amodei wrote in a post earlier this month, “not only investing in risk prevention, but also slowing down the pace of capability development so that risk prevention has time to catch up.” Security training for Opus 5.5 was broadly similar to its predecessors, with alignment testing and pre-assessment by external organizations such as METR and Frontier Design. However, Anthropic stressed that more advanced training and assessment systems are already being prepared for future models, including improved safety and monitoring systems. "As AI becomes
increasingly powerful, public policy must play a greater role in ensuring the security of the systems people rely on. This ability takes time to build, and we have begun laying out the infrastructure to support it,” the blog states. “We expect to share more details about these efforts soon.” The material was prepared by a neural network based on the original publication. AI may make mistakes in details (numbers, names, dates, quotes) - for important information, check with the original source. Anthropic Unveils Claude Opus 5.5 with New Security Measures Delivering Secure High-Performance AI with NVIDIA Confidential Computing Kyutai has released a Voice of Reason model for verbally solving math problems. Nokia opens access to AnyJev, a library for decision-making without training. OpenAI introduced GPT-6 Sol and Luna with new API pricing, reducing the cost by 50%.
