[AINews] GPT 5.6 price cut by 20%-80%: Cost of GPT 5.4 Intelligence dropped 13x in 4 months due to GPT 5.6 recursive self-optimization

One of our big “hero charts” a year ago (eventually adopted by Demis) made the stunning observation that, holding LMSys Elo constant, GPT4 level intelligence fell by 1000x over 18 months:

Back then, it was unclear if these were “noob gains” - going from unoptimized to optimized, going from completions to reasoning, going from dense to MoE, with all the low hanging fruit gone. Another 18 months later, it seems the answer is no; constant-level intelligence is continuing to get precipitously cheaper.

Yesterday, OpenAI published findings on how GPT 5.6 had optimized its own serving:

1. Inference Acceleration

To serve more tokens on the same hardware without sacrificing quality, OpenAI focused on several systemic upgrades:

• Self-Optimization: GPT-5.6 Sol was actively used to analyze production traffic, tune load balancing, and autonomously rewrite production kernels in OpenAI’s own Triton and Gluon languages. This autonomous kernel optimization reduced end-to-end serving costs by 20%.
• Speculative Decoding: Improved its own draft model (by designing and running hundreds of experiments on its architecture, testing changes in size, structure, and features, while monitoring the speculator training process, autonomously intervening when issues arose, including hardware failures and training instability) increasing token-generation efficiency by over 15%.
• KV Caching: Optimized batching, sharding, and cache management specific to different workloads (primarily Sol in Codex) to extract more inference from existing hardware.

2. Agentic Harness Improvements

To streamline the complex, multi-step tasks in tools like Codex and ChatGPT Work, OpenAI optimized their Rust orchestration layer to reduce repetitive compute costs:

• Avoiding Context Bloat: Tools, skills, and plugins are only surfaced when needed (deferred discovery), and tool outputs are capped at 10,000 tokens by default to prevent the context window from expanding unnecessarily.
• Prompt Caching: To avoid reprocessing the same…