GPT 5.6 prices cut 20–80%; intelligence cost drops 13× in 4 months

The article traces a striking pattern of sustained cost reduction in AI inference. A year ago, holding model quality constant (measured by LMSys Elo), GPT-4-level intelligence fell 1000× in cost over 18 months; the new data suggest this deflation has not only continued but accelerated. The specific claim—that GPT-5.4 intelligence now costs roughly one-thirteenth what it did four months earlier—translates to an annualized rate the article frames as "roughly 2000× a year," though it hedges this with the caveat that public benchmarks like Artificial Analysis may be partially trained-to, whereas Elos resist direct optimization.

The price cuts are tied directly to systems-level improvements across three layers: the model itself (self-optimization where GPT-5.6 analyzed production traffic and autonomously rewrote serving kernels), the inference stack (speculative decoding and KV caching tuning), and the agentic harness (context bloat avoidance and prompt caching). This multi-layer optimization appears to reflect a shift from low-hanging fruit (moving from completions to reasoning, from dense to mixture-of-experts) to harder architectural and infrastructure wins. The practical impact is immediate: teams using Codex and ChatGPT are moving auto-review from GPT-5.4 to Luna with an expected 10× cost reduction.

The article also notes that on cost-per-task benchmarks, OpenAI has "beat even open models like DeepSeek, GLM, and MiniMax, and dedicated cheap-but-good models like Gemini Flash-Lite," positioning proprietary inference as the cheapest option for non-finetuned intelligence—a reversal from prior periods when open models offered better unit economics.