Anthropic launches Claude Opus 4.8 with focus on the model being more 'honest' when it makes mistakes

Anthropic, considered a pioneer on artificial intelligence safety, on Thursday released its new flagship model Claude Opus 4.8. According to The Verge, the company has this time pushed forward not the model's technical capability but the training focus that has been invested in the concept of 'honesty.' Anthropic's official statement says: 'We train all of our models to be honest — for instance, to avoid making up claims.'

Among the main innovations in Claude Opus 4.8 is a tool called 'dynamic workflows,' a coordination layer designed for orchestrating sub-agents. According to details carried by The Verge, the model can dynamically change the number of sub-agents and the task distribution over the long output of a task. This delivers significant efficiency gains in applications such as complex code generation, multi-step data analysis and long research assignments.

The technical foundation of the honesty approach rests in the careful rework of the RLHF (reinforcement learning from human feedback) process. Anthropic CEO Dario Amodei, in a press briefing The Verge attended, said: 'We had been working on honesty in our earlier models too, but with Claude Opus 4.8 we applied more strictly a metric known in the AI community as 'relative effort.'' Relative effort measures how well the effort the model applies to a task matches the real complexity of the task.

For example, when a user asks the model for a long research piece, Claude Opus 4.8 actually works for 30-45 minutes depending on the complexity of the task and the output ends up substantially exceeding the quality provided by earlier models. By contrast, if the task is simple, the model gives a quick response while also being able to provide feedback such as 'I don't think this task requires more in-depth research.' According to the content The Verge carries, this approach aims to correct the tendency of earlier models to 'unnecessarily prolong simple tasks with excessive effort.'

In benchmark tests Claude Opus 4.8 achieves significant comparative performance in the sector. According to Anthropic's statements, the model leads OpenAI's GPT-5 (about 72 percent) and Google's Gemini Ultra 3 (about 75 percent) with a 78.4 percent success rate on SWE-bench (software engineering benchmark). On MMLU academic tests it surpassed competitors with a 91.2 percent success rate. These figures aim to reduce barriers to the model's use in areas such as professional software development and academic research.

The pricing strategy is also notable. According to The Verge, the token-based pricing of Claude Opus 4.8 has been set roughly 30 percent above that of its predecessor Claude Opus 4.7. Specifically: 25 dollars per 1 million input tokens; 125 dollars per 1 million output tokens. Anthropic explains the price rise through the 'dynamic workflows' feature and additional investment in the honesty training process. Comparatively OpenAI GPT-5 pricing is 22 dollars per 1 million input; but GPT-5 does not have a feature similar to dynamic workflows.

The safety test process was also pushed forward. Anthropic ran the model through an 18-month red team exercise before releasing it to the public. According to information carried by The Verge, the red team work tested the model's potential side effects such as 'helping in bioweapon production' and 'spreading political misinformation'; these test results were shared with the US AISI (AI Safety Institute). Anthropic CTO Tom Brown said: 'We measured the safety profile of the model as tighter than GPT-5 and Gemini Ultra 3; we will share this in an open report.'

Additional features for enterprise customers also drew attention. Claude Opus 4.8 has an API optimised for production environments and customers can adjust model behaviour for their specific applications through that API. Microsoft, as an Anthropic investor, announced that it would integrate this model into Azure rapidly; AWS is also preparing to offer the model to its users through Bedrock. The Verge says that this enterprise distribution plan will significantly reduce barriers to real-world use of the model.

Reactions from the AI community are mixed. Stanford University AI research director Dr Fei-Fei Li, in comments to The Verge, said: 'Anthropic's honesty approach is an important direction; models acknowledging their own limits is a cornerstone of AI safety.' By contrast, former OpenAI chief scientist Ilya Sutskever wrote on X (Twitter): 'Claiming that models are honest is a concept hard to measure; the real honesty test is set by how the model behaves in real-world use.'

Taken in the round, the release of Claude Opus 4.8 can be read as a milestone showing how much AI models focus, beyond capabilities, on the area of 'behavioural quality.' According to The Verge's final word, the public production experience of the model will be shaped over the coming weeks by user feedback. Anthropic says the model will continue to evolve and plans to expand the model's capabilities using new training data over the next three months. This article is not investment or technology procurement advice; the technical data rest on Anthropic's official statement and The Verge's test reports.