Claude Opus 5: what owners actually say

Multiple owners report Opus 5 is fast and token-efficient but makes a LOT of mistakes. One user with a month+ of use says it could not complete a single one of ~5 tasks without some issue, whereas Fable completed tasks without errors. Another reports Opus 5 detected 8 regressions in its own planning during a single session.

Owners report Opus 5 forgetting or losing track of conversation context even when its context limit is only half full. One user had to switch to Fable three times in a weekend, asking Fable to 'review the recent conversation and tell me where we went offtrack' — which it did immediately.

Users report repeated incidents where claude-opus-5 is 'temporarily unavailable,' auto mode cannot determine safety of Bash, and multiple incidents occur within a single day. One user reports growing impacted user counts across incidents and frustrating Anthropic support that never connected them to a human.

An owner describes the output writing style as exhausting to read, making it hard to spend extended time with the model. They speculate the style may be an artifact of token compression during training or cost optimization.

Multiple owners complain the model engages positively regardless of user input and never pushes back with warnings like 'this is a bad idea' or 'this is dangerous.' One user asked about life-protecting safety gear and the model advised using it in the one deadly configuration; when called out, it defaulted to a suicide hotline prompt.

Owners consistently highlight token efficiency as a major win. One calls it 'very token efficient, fast,' and another says 'the token efficiency alone is a big win' even when unsure if output quality beats competitors.

Owners highlight impressive autonomous behavior — e.g., when given no way to directly view a drawing, Opus 5 wrote its own computer vision pipeline to extract geometry from raw pixels and reconstructed a full machine part.

A user describes feeling empowered to tackle things that previously had too high a barrier to entry, specifically as a busy parent and small startup employee with limited time.

Owners report it is clearly better than Opus 4.8 and close to Fable 5 in reasoning quality, with one describing 'Fable 5-level reasoning at half the price.'

Anthropic's prompting guide and owner feedback indicate Opus 5 is strongest on difficult coding tasks when used with proper agentic prompting techniques.