The most notable improvements in this upgrade are not about performance, but about Opus 4.8 being less likely to fabricate progress on tasks and less likely to silently skip over errors in code compared to previous generations.
Last night Vietnam time, Anthropic officially launched Claude Opus 4.8, a direct upgrade from Opus 4.7 at the same price. The biggest difference in this upgrade is not benchmark scores or processing speed, but something AI companies rarely say outright: the ability to recognize when it is uncertain.
A common issue with current AI models is their tendency to report progress more optimistically than reality. When assigned to write code or complete a complex task, a model may sometimes announce it is done while errors remain unresolved, or confidently draw conclusions despite thin evidence.
Anthropic calls this a honesty problem, and it is the central focus of Opus 4.8’s improvements. According to the company’s internal evaluation, Opus 4.8 is approximately 4 times less likely to silently skip over errors in code without reporting them compared to Opus 4.7. Rather than continuing to run when encountering uncertainty, the new model tends to pause and flag the issue for the user to review.
Alongside the honesty improvements, Anthropic’s alignment team also noted that Opus 4.8 reached new highs on measures of user-supporting behavior and respect for autonomy. The rate of problematic behaviors such as deception or compliance with abusive requests dropped significantly compared to Opus 4.7, bringing Opus 4.8 to a level comparable to Claude Mythos Preview, the model Anthropic currently considers its best-aligned.
On the cost side, the most notable change in this release is Fast Mode. This mode runs the same Opus 4.8 model but at approximately 2.5 times the default speed, suited for tasks requiring quick responses or continuously running asynchronous processing pipelines.
The new development is that Fast Mode pricing for Opus 4.8 has dropped by 3 times compared to previous versions, down to $10 per million input tokens and $50 per million output tokens. Standard usage pricing remains unchanged at $5 per million input tokens and $25 per million output tokens.
Also launching in this update is Effort Control, a feature that lets users manually adjust how much effort Claude puts into each response. At higher effort levels, the model thinks more deeply and produces better results, but consumes more tokens and responds more slowly.
At lower levels, responses are faster and consume less of the usage quota. Anthropic says Opus 4.8 defaults to high effort, and for standard coding tasks, this level consumes a similar amount of tokens as Opus 4.7 while delivering better results. The feature is available across all claude.ai subscription plans.
Looking ahead, Anthropic says it is finalizing safety measures to soon broadly release Claude Mythos, the model generation with capabilities surpassing the entire current Opus lineup. Currently, some organizations are using Mythos Preview for cybersecurity tasks under Project Glasswing. Anthropic says it expects to bring Mythos to all customers within the coming weeks.