OpenAI's GPT-5.6 family ships in three sizes, each with five or six reasoning-effort settings. That design choice is not cosmetic. It reflects two years of industry convergence on a single idea: reasoning compute is a dial, not a switch. OpenAI's o1 started it. DeepSeek-R1 followed four months later with a public reinforcement learning with verifiable rewards (RLVR) recipe, and now variable-effort reasoning is a baseline expectation for any serious model release.
The original piece traces that full arc, from the RLVR training methodology behind DeepSeek-R1 to how GPT-5.6's Ultra tier deploys four parallel subagents to hit similar benchmark ceilings as its Max setting. The benchmark comparison across effort levels is the most useful section: it shows exactly where additional compute stops paying off, and that threshold differs by task type. That data is the reason to open the full article.
Author Sebastian Raschka has also released a 440-page book titled 'Build A Reasoning Model From Scratch,' which makes this piece partly a roadmap for that deeper material. If you work on inference budgets, agent orchestration, or LLM evaluation, the effort-versus-performance curves documented here will inform decisions you are already making.
[READ ORIGINAL →]