Usually, when Anthropic announces a new model, I talk about how it is smarter, more token-friendly, or how much it outperforms the previous one on benchmarks. That is still the case with Claude Opus 4.8 but there is something else more interesting about it that I would rather get into. It has always been annoying to me when an AI chatbot has confidently continued to give answers when it is clear as day that it has no idea, Opus 4.8 might just be changing that. According to Anthropic, early testers have said that “Opus 4.8 is more likely to flag uncertainties about its work and less likely to make unsupported claims.” This is something that we are still yet to see but if it is true, it is a much bigger story than any benchmark.
Also read: “Don’t build around one processor”: DFI President Smit Shah India’s drone supply chain
Numbers back this up to some extent. Anthropic reports that Opus 4.8 is about four times less likely than its predecessor to fail to flag flaws in the code it has generated. That’s no small accomplishment. Anyone who has dug through the code that an AI was absolutely certain of – only to find it plagued with subtle errors – knows the significance of such a metric all too well. Confidence that is unfounded is not an asset, it’s a liability.
The honesty issue in AI is not merely a technical one. What Anthropic describes here is a system that is aware of its limitations – that doesn’t overpromise but instead tries to deliver on its promises. The entire journey of artificial intelligence has taken us in the direction of an appearance of certainty – perhaps, because it’s what we have been seeking all along. But what good is certainty when you know next to nothing? An answer hedged in uncertainty may seem like a less confident response. However, the opposite just might be true.
Also read: What happens to our brain in deep sleep? AI finally has an answer
Opus 4.8 also remains an improvement over its predecessor in coding, reasoning, and agentic tasks. Furthermore, Opus 4.8 retains the price of Opus 4.7, which is $5 per million input tokens and $25 per million output tokens. Anthropic has also introduced a “fast mode,” which operates twice as fast and costs three times less than its equivalent cost for the previous generation of models. Moreover, control over effort has also been enabled on claude.ai, and users can now increase or decrease effort spent by the model on particular tasks.
None of that, though, is the story. The story is a model that, when it does not know something, says so. In a space that has been drowning in confident nonsense, that is the most interesting thing Anthropic has shipped in a while.
Also read: Don’t want AI in your search? Try these 5 traditional search engines