Claude Opus 4.8 is an incremental upgrade to Opus 4.7 with improved honesty and fewer unnoticed coding errors. The model flags uncertainties more quickly and makes fewer unsupported claims about its own work. Alignment assessments show higher prosocial traits, including support for user autonomy and acting in the user’s best interest. Practical performance improves across agentic coding, multidisciplinary reasoning, computer use, knowledge work, and financial analysis, with gains ranging from under 1 percentage point to nearly 9 percent. Deception and misuse-related cooperation rates are substantially lower than Opus 4.7 and are comparable to Claude Mythos Preview under restricted access. Pricing remains $5 per million input tokens and $25 per million output tokens, alongside platform updates such as dynamic workflows and user-controlled effort levels.
"The new model is more honest in its self-assessments and roughly four times less likely to let coding errors pass unremarked. It launches at the same price as its predecessor, alongside new features including dynamic workflows and user-controlled effort levels. Despite assumptions that Anthropic will eventually release a Mythos-level LLM, this doesn't appear to be anywhere close to that presumed capability. The improvements in honesty are what sets Opus 4.8 apart from its direct predecessor, according to Anthropic."
"Early testers reported the model is quicker to flag uncertainties and less likely to make unsupported claims about its own work. According to Anthropic's alignment assessment, Opus 4.8 "reaches new highs on our measures of prosocial traits like supporting user autonomy and acting in the user's best interest." In terms of practical benefits, Opus 4.8 reaches new heights in agentic coding, multidisciplinary reasoning, computer use, knowledge work and financial analysis."
"Improvements range from less than 1 percentage point to nearly 9 percent. The difference between 4.7 and 4.8 in these statistics suggest the informal, unmeasured day-to-day experience won't be all that different at any given point, while still boosting overall results long-term. Time will tell if that is true. That aforementioned alignment assessment showed rates of deception or cooperation with misuse being substantially lower than those of Opus 4.7."
"Notably, those rates are now comparable to Claude Mythos Preview, the powerful model Anthropic has kept under tight access restrictions due to its advanced cybersecurity capabilities. Pricing is unchanged for Opus, at any rate, at $5 per million input tokens and $25 per million output tokens. Dynamic workflows and effort control The launch comes bundled with several platform updates. The most significant for developers is dynamic workflows, currently available"
Read at Techzine Global
Unable to calculate read time
Collection
[
|
...
]