FULL STORY

GPT-5.6: From Self-Optimization to Surging Benchmarks

OpenAI revealed that GPT-5.6 optimized its own inference stack to cut costs, while developers discovered that enabling specific API settings triples its score on the ARC-AGI-3 benchmark.

2026-07-30 ~ 2026-07-30 · 2 episodes · 25 posts

Episode 1 · OpenAI Says GPT-5.6 Sol Self-Optimizes, Cutting Inference Costs by 20% (2026-07-30, 8 posts)

OpenAI announced that following the deployment of the GPT-5.6 (Sol) model, it utilized the model to achieve self-optimization of its own inference stack, successfully merging frontier intelligence with operational efficiency. Official data shows this optimization reduced service costs by 20% and boosted Token generation efficiency by over 15%. This signals that large models are now capable of optimizing underlying infrastructure to reduce steep inference costs, a crucial development for the industry.

Confirmed

  • Model and Deployment: OpenAI has deployed the GPT-5.6 Sol model, running and testing optimizations within the Codex environment.
  • Service Cost Reduction: By improving GPU kernels in production environments (rewriting low-level code), GPT-5.6 Sol successfully reduced end-to-end service costs by 20%.
  • Generation Efficiency Boost: Through improved speculative decoding (and helping optimize smaller models used to accelerate generation), Token generation efficiency increased by over 15%.

Why It Matters

  • AI Self-Optimization: This wasn't purely manual tuning; the GPT-5.6 Sol model directly participated in enhancing its own operational efficiency. This demonstrates that advanced AI models can not only execute complex tasks but also act as optimization tools, directly solving the industry pain point of high inference costs, and paving a new path for commercialization and operational efficiency.

Episode 2 · GPT-5.6 Scores Triple on ARC-AGI-3 After Enabling Two API Settings (2026-07-30, 17 posts)

GPT-5.6's performance on the ARC-AGI-3 benchmark varies drastically depending on API configurations. Developer tests reveal that enabling two specific settings used internally by ChatGPT and Codex boosts the model's score by roughly 3x to 38%, while improving token efficiency by about 6x. This highlights the critical synergy between model capabilities and product frameworks, suggesting that isolated evaluations may severely underestimate a model's true potential.

已确认

  • Developer @sandersted found that GPT-5.6 performs poorly by default on the ARC-AGI-3 benchmark
  • Enabling two specific API settings used internally by ChatGPT and Codex increased the public test set score by about 3x, reaching 38%
  • Token efficiency simultaneously improved by roughly 6x
  • These two settings are 上下文压缩（compaction） and 推理记忆保留（reasoning retention）, allowing the model to remember previous thoughts
  • OpenAI engineer @ilanbigio pointed out that model evaluations should generally not be conducted in isolated environments

为什么重要

  • This discovery challenges the industry's current isolated evaluation methods, suggesting they may fail to reflect a model's actual capabilities in real-world product environments
  • The deep synergy between product frameworks (such as ChatGPT and Codex's internal mechanisms) and model capabilities has a decisive impact on performance in complex tasks
  • This also implies that simply comparing scores from bare-bones benchmark runs may not accurately measure the true intelligence of different AI systems