Reasoning
GPT-4-class input pricing fell from $30/M tokens in March 2023 to $2.50/M by July 2025 (91.7% drop), driven by 4× inference efficiency gains per generation and hyperscaler capex (Microsoft $50B, Google $48B in FY2025). Current 2026 H1 spot pricing remains at $2.50/M with no further announced cuts, yet next-gen models (GPT-5-class) are slated for 2026H2 release and have already shown 3–4× training FLOP efficiency in internal benchmarks; combined with continued 30–40% annual hardware cost declines, a further 75%+ reduction by Q4 2026 remains the base case.Key uncertainty
Whether OpenAI or competitors delay price cuts to protect margins amid rising inference demand from agentic workloads.