Claude's Fable 5 shows diminishing returns in real-world software engineering

A distinguished engineer at a hyperscaler recently got hands-on with Claude's latest Fable 5 model and came away unimpressed. In blind tests comparing Fable 5 with Opus 4.6, 4.7, and 4.8, he couldn't tell which model he was using for his typical software engineering tasks. His reasoning: he never one-shots projects. Instead, he works in small, incremental chunks — testing single abstractions before moving on. Since models already have access to the internet's full suite of documentation and best practices, the added 'intelligence' of newer models provides little marginal gain for this workflow.

The engineer also noted that Fable 5 still hallucinates on complex system architecture. For example, it got AWS ALB/ECS draining behavior completely wrong, and with high confidence — only his domain expertise caught the error. He believes the industry is hitting an asymptotic limit where each new model release offers diminishing returns for hands-on software engineers. Looking ahead, he predicts that within a year, local models running on a 128GB MacBook Pro will deliver 90% of the value Claude provides today, already visible in current open-source models like Gemma 4.