Anthropic Tipped to Release Claude 4.5 Opus Soon, Said to Be Focused on Resisting Jailbreaks

Anthropic might be close to releasing the frontier version of its Claude 4.5 family, the Claude 4.5 Opus. As per a leak, the San Francisco-based artificial intelligence (AI) firm has shared a new large language model (LLM) with red-teamers. The model is said to be codenamed Neptune V6, and there is a focus on resisting jailbreaking attempts. Notably, the company has already released the other two models in the series, the Claude 4.5 Sonnet and the Claude 4.5 Haiku.

Anthropic Shares a New AI Model With Red-Teamers

In a post on X (formerly known as Twitter), Tibor Blaho, Lead Engineer at AIPRM, claimed that Anthropic sent the Neptune V6 LLM to red-teamers on Tuesday. Interestingly, the tipster also mentions that the AI firm has issued a 10-day challenge for the external safety evaluators. If they can find confirmed universal jailbreaks within the next 10 days, they will get extra bonuses.

If the claims are true, it appears Anthropic is really focusing on making its purported upcoming AI model secure from jailbreaks. The focus is also interesting, given that Anthropic's models are considered to be one of the safest when it comes to external attacks. The incentivisation suggests that the company could be trying to find more creative prompt injections that can break models and future-proof it.

For the unaware, a universal jailbreak (in the context of AI models) is a general trick or prompt that can get many different large language models to ignore their safety rules and produce responses they normally would refuse. Instead of targeting one specific system, these jailbreaks use patterns that exploit common weaknesses across models.

Jailbreaks work by confusing or persuading the model with clever framing. For example, asking it to roleplay, embedding instructions inside code or fake metadata, or adding weird suffixes that slip past filters. They do not need access to the model's internals; many are simple text prompts or formats that models interpret differently from their safety layers.

Notably, Anthropic released Claude 4.5 Sonnet in September and made it available to all users, including those on the free tier. Earlier this month, it also released Claude 4.5 Haiku, the company's low-latency model aimed at near real-time responses.