GPT-5.6 Price Cut 20–80%: Cost of GPT-5.4 Intelligence Dropped 13x in 4 Months via 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 instructions and history repeatedly, the harness treats all model-visible history as append-only. This preserves the prompt prefix, allowing the system to reuse previously computed data and maintain a high cache hit rate. Today, they proved it wasn’t just theory, announcing large price cuts for the smaller models and a new 2.5x Faster mode in Sol (though not the hinted 10x faster Cerebras-driven mode promised in July): That is a beautiful Pareto curve — if AA’s Cost per Task is reflective of real world tasks, then OpenAI has beat even open models like DeepSeek, GLM, and MiniMax, and dedicated cheap-but-good models like Gemini Flash-Lite. Good, but what is truly jaw dropping is Nicdunz’s observation : GPT-5.4 full at xhigh scored 51, exactly where Luna max sits today . GPT-5.4 costs $2.50/$15; Luna now costs $0.20/$1.20. In other words, roughly four months later, OpenAI is selling March’s full flagship intelligence at about one-thirteenth the token price . That’s an annualized rate of ~2000x a year, a huge acceleration since our last observation - which you SHOULD discount because all public benchmarks like AA’s get trained to some extent whereas Elos are less directly trainable. While leading Chinese and American open models like Poolside’s Laguna and Thinky’s Inkling offer full control and sovereignty, if your only goal is cost-effective non-finetuned intelligence, you will find it -very- hard to beat OpenAI right now. AI News for 7/29/2026-7/30/2026. We checked 12 subreddits, 544 Twitters and no further Discords. AINews’ website lets you search all past issues. As a reminder, AINews is now a section of Latent Space . You can opt in/out of email frequencies! AI Twitter Recap OpenAI Pricing Cuts, Harness Semantics, and the ARC-AGI-3 Memory Debate OpenAI cut GPT-5.6 prices aggressively and added a faster Sol tier : OpenAI reduced GPT-5.6 Luna by 80% and Terra by 20% , while introducing Sol Fast at up to 2.5× lower latency for 2× the standard price with “no change in intelligence,” per @OpenAIDevs . The downstream effect is notable for agent workflows: Auto-review in ChatGPT app and Codex CLI is moving from GPT-5.4 to Luna , with OpenAI expecting roughly 10× lower cost . Multiple observers framed this as a meaningful shift in the price/performance frontier, including @sama , @nicdunz , and @kimmonismus . OpenAI also tied the cuts to systems-level efficiency improvements spanning “model, inference stack, and agentic harness” per @OpenAIDevs . ARC-AGI-3 re-emphasized that “the model” is not the whole system : The most technical eval discussion centered on harness design, memory retention, and context compaction. François Chollet clarified ARC’s rules: bespoke benchmark-specific harnesses are disallowed, but general-purpose API features available to all users are acceptable if settings and cost are reported. A detailed summary from @kimmonismus contrasted Opus 5 at 30.2% on the official semi-private ARC setup with GPT-5.6 Sol at 7.8% under the standard harness, while noting OpenAI’s internal use of Responses API retained reasoning + compaction raised Sol’s public-set score to 38.3% . The takeaway echoed by @gneubig , @scaling01 , and others: long-horizon evals increasingly measure the complete agent system —reasoning retention, truncation policy, compaction, tool orchestration—not just base weights. Thinking Machines’ Inkling-Small and the Continuing Open-Weights Push Inkling-Small compresses Inkling-class capability into a much smaller active footprint : Thinking Machines released Inkling-Small , an open-weights , natively multimodal MoE model with 276B total parameters and 12B active , positioned as delivering performance comparable to the original Inkling at roughly a quarter the size. The company says it processes audio and images jointly with text and supports Python-based image inspection during reasoning in multimodal settings, per their follow-up . The release immediately landed across the open inference stack: vLLM announced day-0 support , Modal highlighted single-B300 deployment , LMSYS/SGLang reported decode throughput figures , and Unsloth published a local-running/GGUF guide . Benchmarks suggest an unusually efficient open model for coding and multimodality : Artificial Analysis placed Inkling-Small at 40 on its Intelligence Index—within a point of the flagship Inkling—with strengths on Humanity’s Last Exam, GPQA Diamond, CritPt, and SciCode , though weaker on some agentic tasks and factual knowledge. Community summaries emphasized that the smaller model can beat or match the larger Inkling on several coding tasks , including @kimmonismus and @mervenoyann . The combination of open weights, multimodal input, 1M-context support in deployment stacks, and 12B active compute makes this one of the more practically important open releases in the batch. Google’s Gemini Robotics 2 and the Acceleration of Embodied AI Gemini Robotics 2 expands from tabletop manipulation to full-body control and multi-robot coordination : Google DeepMind launched Gemini Robotics 2 , describing it as “one brain for any robot,” with demos spanning whole-body humanoid control , advanced dexterity , and multi-robot collaboration . Google AI added that the stack includes Gemini Robotics ER 2 , a high-level embodied reasoning model that can observe, plan, coordinate with a VLA model, track progress, and recover from failed steps during multi-minute tasks. The demos emphasized nontrivial motor tasks such as knot-tying, screwing in a bulb, bending to pick up objects, and collaborative garage cleanup. The practical story is heterogeneity and adaptation, not just nicer demos : Technical commentary highlighted that the same checkpoint controlled multiple hardware types and that On-Device 2 can reported