OpenAI GPT-5.6: Smarter, Cheaper

OpenAI has unveiled the GPT-5.6 model family, touting substantial gains in both computational efficiency and AI capability. The new lineup aims to balance performance with cost, a critical factor as AI demand outstrips supply. The flagship model, GPT-5.6 Sol, claims superior reasoning over competitors at less than half the price. Terra matches GPT-5.5's intelligence at half the cost, while Luna offers the fastest and most affordable performance, priced 80% lower than Sol. These advancements stem from deep optimizations across OpenAI's stack, impacting models, inference processes, and their agentic harness used in products like Codex and ChatGPT Work.

The company emphasized that efficiency has been central to its mission of democratizing AI. OpenAI reports achieving its highest intelligence-per-token efficiency yet with GPT-5.6, which is trained to perform more work with fewer tokens. This training focuses on both task success and efficiency, guiding models to find more direct solutions. This post details how OpenAI achieved these gains not just through model improvements but also through advancements in inference and the agentic harness.

Accelerating Inference

Inference, the process of running trained models to generate output, is a primary target for efficiency gains. OpenAI’s goal is to serve more tokens using the same hardware while maintaining quality, latency, and reliability. Improvements compound across routing, scheduling, kernel optimization, caching, and model implementation.

GPT-5.6 Sol played a direct role in optimizing these areas. Load balancing improvements alone significantly reduced serving costs by analyzing production traffic and tuning routing heuristics globally and within clusters. StartupHub.ai data shows OpenAI itself scores an 84/100 in its competitive landscape, suggesting a strong internal focus on such efficiencies. Codex, powered by GPT-5.6 Sol, autonomously rewrote and optimized production kernels, reducing end-to-end serving costs by 20%. These kernels are written in Triton and Gluon, open-source GPU programming languages maintained by OpenAI.

Speculative decoding, a technique using a smaller draft model to predict tokens for parallel verification by the main model, saw a 15% increase in token-generation efficiency. GPT-5.6 Sol improved its own draft model through extensive experimentation and monitored the training process autonomously. Furthermore, GPT-5.6 Sol in Codex enabled hyper-optimization of inference configurations based on workload analysis, extracting more performance from existing hardware. This iterative process of measuring, identifying gaps, implementing changes, and verifying results accelerates development.

Streamlining the Agentic Harness

The agentic harness, an orchestration layer connecting models, tools, and user environments, was also a focus. This system handles complex tasks involving multiple model requests and tool calls. Reducing repeated work within these multi-step processes is key to overall performance.

To combat context bloat, the harness employs deferred discovery, making integrations and tools available only when needed. It also caps tool output at 10,000 tokens by default. Prompt caching was enhanced by treating model-visible history as append-only, preventing costly recomputation of prompt prefixes. Tools are presented deterministically, and runtime settings are applied during execution rather than embedded in definitions. These changes contribute to high prompt-cache hit rates for products like Codex and ChatGPT Work.

The efficiency gains from GPT-5.6 represent years of compounding improvements across research, inference, and the agentic harness. OpenAI anticipates this pace of optimization will accelerate, leading to more accessible and cost-effective AI for users and customers.