Inside GPT‑5.6’s Dropdown: How Six Leading LLMs Tune Their Reasoning Effort

1. Reasoning Effort Becomes a Flagship Feature

OpenAI’s GPT‑5.6 series (Luna, Terra, Sol) ships with five to six predefined effort levels (Light → Medium → High → Extra High → Max → Ultra). The author shows a cost‑performance curve where the Artificial Analysis Coding Agent Index rises from ~58 to ~81 points while API cost climbs from a few hundred to over 2000 USD, with diminishing returns at higher levels.

2. What a “reasoning model” actually is

Raschka stresses the old principle “don’t take technical terms literally”. A reasoning model does not think like a human; it merely emits a step‑by‑step reasoning trace, enabling self‑correction and back‑tracking, unlike traditional LLMs that produce a short answer without an inspectable process.

3. Two Scaling Axes

Improving reasoning capability can follow two orthogonal routes:

Training scaling : invest more compute during training (larger models, more RL).

Inference scaling : invest more compute at inference time (longer chain‑of‑thought, multiple sampling).

4. RLVR Training Recipe (Three Key Points)

(1) Reward only result and format. DeepSeek‑R1’s RLVR uses a binary 0/1 reward based on verifiable domains (e.g., SymPy for math, unit tests for code): R_total = R_accuracy + R_format.

(2) Reasoning trace is not used for training. Experiments showed that feeding intermediate traces as training signals did not help; the process‑reward model (PRM) remains an active research direction but was dropped for the final RLVR.

(3) Pure result reward yields an “Aha moment”. R1‑Zero demonstrated that pure RL (no SFT) can spontaneously produce reasoning behavior, writing intermediate explanations, spotting errors, and self‑correcting.

5. The <think> Tag Is Cosmetic

The <think></think> token pair only marks the start and end of a reasoning block for UI and training pipelines; it does not endow the model with thinking ability. Any other delimiter works, and models can be trained without it while achieving comparable benchmark performance.

6. First‑Generation On/Off Switch: Qwen3’s Thinking Mode Fusion

Early reasoning models (e.g., R1) always produced long chains of thought. Qwen3 introduced a mixed‑mode approach: during SFT, samples are labeled /think (with reasoning trace) or /no_think (short answer). At inference, the soft switch ( /think vs. /no_think) and a hard switch ( enable_thinking=False or explicit <think></think>) control whether the model emits a reasoning block.

7. How Effort Levels Are Implemented

OpenAI does not disclose details, but the open‑source gpt‑oss repository reveals that effort is injected via a system prompt (e.g., “Reasoning effort: low”). The prompt influences token budget, length, and accuracy, effectively acting as a “token‑budget knob”.

8. Typical Training Implementations

Effort‑conditioned RLVR : during RL, the system‑prompt effort tag determines a length penalty (low effort → high penalty, high effort → low penalty).

Effort‑conditioned SFT : after RLVR, samples pair an “effort‑label” with a target length, teaching the model the relationship directly.

Both approaches can be combined; gpt‑oss and GPT‑5.6 likely use a hybrid.

9. Shared Components Across the Six Flagship Models

DeepSeek V4 – three modes (Non‑think / Think High / Think Max) with distinct context windows and length penalties; Think Max gets a special system instruction.

Nemotron 3 Ultra – learned mode + hard budget dual‑track; medium‑effort uses GPT‑OSS‑120B trajectories; random truncation of reasoning traces forces the model to answer directly.

Kimi K2.5 – “Toggle” method alternating between budget‑constrained and unconstrained RL, saving 25‑30 % tokens with negligible performance loss, transferable to GPQA/MMLU‑Pro.

GLM‑5 – multiple “thinking” switches per turn, interleaved thinking before tool calls, and preserved thinking across turns, all introduced via multi‑task SFT and a new chat template.

Qwen3 – Mode Fusion (mix of think/no‑think samples) plus hard inference‑time budget cuts; the ability to continue answering after truncation emerged spontaneously.

Inkling – continuous effort value (0.0‑1.0) conditioned RL; reward formula R = R_accuracy − λ(e)·n_tokens adjusts token cost per effort, encouraging short reasoning at low effort and longer chains at high effort.

10. Short‑Term Outlook

Effort will stay an explicit system‑prompt input for now, but routing/harness components are expected to infer the appropriate mode and token budget automatically based on task state and resource availability, while still allowing user overrides.