The Benchmark Treadmill and What Actually Matters
GPT-5 launches with the now-familiar ceremony: benchmark scores that beat every previous model, capability demos that feel genuinely impressive in isolation, and a pricing structure that rewards high-volume enterprise users. The benchmarks matter less than commentary suggests. Every lab has learned to optimise for benchmark performance in ways that do not always translate to real-world usefulness. What matters is whether the model is actually better at the things you try to use it for every day. The reasoning improvements are real. GPT-5 handles multi-step problems with significantly less hallucination than its predecessor. The context window is longer and the model degrades more gracefully at long contexts than GPT-4 did. The multimodal capabilities are also meaningfully upgraded — image understanding in particular feels qualitatively different, as the model can reason about visual content rather than simply describe it.
OpenAI is releasing GPT-5 into a market that looks nothing like the one GPT-4 arrived in. Anthropic's Claude, Google's Gemini Ultra, and Meta's Llama series have all closed the capability gap substantially. The frontier is now crowded. OpenAI's moat, if it still exists, is distribution — ChatGPT has over 180 million weekly active users. Model capability alone no longer differentiates the products that most users encounter, which is why the competition has shifted from research benchmarks to product experiences, integration ecosystems, and enterprise relationships.
What the Next 18 Months Will Determine
The most significant near-term competition in foundation models is not for best benchmark performance but for best reasoning performance on multi-step real-world tasks, best performance in agentic settings where the model must take actions rather than just produce text, and best integration with enterprise software environments that already store the context the model needs to be useful. These are the dimensions on which GPT-5 will actually be judged by the organisations deploying it at scale, and they are dimensions on which the published benchmarks provide only indirect evidence. The 18 months following any major model release have historically produced more information about competitive positioning than the release itself.