GPT-4.5 review: improved verbal intelligence, contextual adaptation, abstract writing, and better taste and aesthetics, but not ideal for reasoning or coding
Zvi Mowshowitz / Don't Worry About the Vase :
Context & Ripple Effects
GPT-4.5 sits in an arc that began with GPT-4’s gains in precision and accuracy, even as that earlier model continued to hallucinate. This review distinguishes progress in language-facing capabilities from progress on the more demanding work of reasoning and code generation.
The distinction matters because later coverage finds that choosing the right model still matters, rather than treating successive releases as universal replacements. GPT-4.5’s value in this record is therefore task-specific: verbal and stylistic work, not a blanket performance upgrade.
First-order effects
- Users doing abstract writing or work requiring adaptation to context gain a stronger option, with the review also crediting GPT-4.5’s aesthetic judgment.
- Users whose primary needs are reasoning or coding have a clear reason not to make GPT-4.5 their default model.
Second-order effects
- Teams adopting OpenAI models will need to route work by task type, separating writing-oriented prompts from reasoning and software tasks rather than standardizing on one model.
- Model evaluations will put greater weight on qualitative attributes such as tone, taste, and contextual fit—not only benchmark-style accuracy—while retaining separate tests for coding and reasoning.
Third-order effects
- If this split persists, frontier-model competition may increasingly center on portfolios of specialized capabilities and routing interfaces rather than a single model that is best at every task.
- The economic value of model access will depend more on matching capability to workload, an emerging form of technical reasoning and coding differentiation that complicates simple “newest model wins” purchasing decisions.
The trend: This is one data point in the shift from measuring generative AI as a single general capability toward selecting models according to distinct writing, reasoning, and coding strengths.