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

On July 30, 2026, OpenAI announced steep price cuts for GPT-5.6 models and a faster inference mode, the culmination of months of backend optimization. Sam Altman posted the headline numbers: GPT-5.6 Luna dropped 80% to $0.20 per million input tokens and $1.20 per million output tokens; Terra fell 20% to $2/$12; and Sol gained a Fast mode offering up to 2.5× speed for 2× the standard price, with "no change in intelligence."

The cuts emerged from a detailed systems optimization effort spanning three domains. For self-optimization, OpenAI used GPT-5.6 Sol itself to analyze production traffic, tune load balancing, and autonomously rewrite production kernels in Triton and Gluon, reducing end-to-end serving costs by 20%. On speculative decoding—a technique where a smaller model drafts tokens and a larger model validates them—GPT-5.6 designed and ran hundreds of experiments, testing changes in size, structure, and features while monitoring training and autonomously intervening during hardware failures and training instability, increasing token-generation efficiency by over 15%. For KV caching (a memory optimization for inference), the team optimized batching, sharding, and cache management for different workloads, especially Sol in Codex. To streamline multi-step agentic tasks, OpenAI optimized the Rust orchestration layer, implementing deferred tool discovery (tools surface only when needed) and capping tool outputs at 10,000 tokens to prevent context bloat. Prompt caching treats all model-visible history as append-only, preserving the prompt prefix and allowing the system to reuse previously computed data at high cache-hit rates.

The market impact was sharp. An observer noted that GPT-5.4 full at "xhigh" quality scored 51 on Artificial Analysis Cost per Task, exactly where Luna max now sits. GPT-5.4 cost $2.50/$15; Luna now costs $0.20/$1.20—roughly one-thirteenth the token price, achieved in approximately four months. The article frames this as an annualized rate of roughly 2000× cost reduction per year, with the caveat that public benchmarks may be partially trained-to. The downstream effect is immediate: auto-review in ChatGPT and Codex CLI is migrating from GPT-5.4 to Luna, with OpenAI expecting roughly 10× lower cost. The article notes that on cost-per-task measures, OpenAI has surpassed open models like DeepSeek, GLM, and MiniMax, as well as specialized cheap-but-good models like Gemini Flash-Lite, making proprietary non-finetuned inference the lowest-cost option available.