"LLM Reasoning Modes (6/6) — Comparison and Practical Guide: Cost, Latency, Quality, and Which Dial When"

We've taken the dial apart piece by piece — Part 1 the principle, Part 2 Claude's thinking, Parts 3-4 Claude effort, Part 5 OpenAI reasoning_effort. Part 6 puts the two side by side and answers, with benchmarks, "does more thinking always help?" Then it tabulates which notch to use per task.

The one-line conclusion of this series, up front: the reasoning dial is both a quality knob and a cost knob, and the right answer is not "always maximum" but "match it to task difficulty." Part 6 gathers the comparison, figures, and guide that back that up.

In One Paragraph

Claude's effort controls thinking, text, and tool calls with a single dial (default high); OpenAI splits thinking depth (reasoning_effort, default medium) and output length (verbosity) into two dials. Benchmarks show the dial is a cost-efficiency lever — per Anthropic, on SWE-bench Verified, Opus 4.5 at medium effort matched Sonnet 4.5's best score using 76% fewer output tokens, and at high exceeded it by 4.3 points using 48% fewer tokens. More thinking is not always better (overthinking). Practical rule: low for simple tasks, high for hard reasoning/agents, top notch only for genuine frontier problems.

1. Claude vs OpenAI — One-Table Comparison

Both platforms address the same tradeoff (quality ↔ cost/latency) with different design philosophies.

Dimension Claude effort OpenAI reasoning_effort
Levels low · medium · high · xhigh · max none·minimal · low · medium · high · xhigh (varies by model)
Default high medium
Scope all tokens (thinking + text + tool calls) thinking depth. Output length is separate (verbosity)
Thinking control adaptive thinking + effort (one dial) reasoning_effort + verbosity (two dials)
Lowest notch low (still thinks on hard problems) none / minimal (≈ no reasoning)
Raw chain of thought summary only, never returned verbatim hidden reasoning + optional summary
Loop-cost knob Task Budgets (countdown, min 20k) (none separate — effort + output cap)

Two key differences:

  • Single vs split. Claude's one dial moves thinking depth and output/tool verbosity together. OpenAI turns thinking (reasoning_effort) and output length (verbosity) separately. "Think deeply but answer briefly" is two explicit parameters on OpenAI; on Claude you supplement with the prompt.
  • Defaults. Claude high, OpenAI medium. The same "leave it at default" means different amounts of thinking, so cost and latency feel different when you move between platforms.

2. Cost, Latency, Quality — Three Variables That Move Together

Recall the curve from Part 1. Let test-time compute be \(C_{\text{test}}\); on hard problems accuracy generally rises as \(C_{\text{test}}\) grows but with diminishing returns. Turning the dial up means growing \(C_{\text{test}}\), and three variables move with it.

  • Quality ↑ — but only on hard tasks, and increasingly gently.
  • Cost ↑ — thinking/reasoning tokens are billed at the expensive output-token rate (both platforms). Turn the dial up and even the invisible tokens grow, inflating the bill.
  • Latency ↑ — thinking must finish before the first answer token, so time-to-first-token rises.

So the dial isn't a pure "be smarter" button — it's a knob over how much cost and latency you'll pay for quality. In the flat region of the curve, turning it higher barely moves quality while cost and latency keep climbing — that flat region is the overthinking of the next section.

3. What the Benchmarks Say — Smarter Model + Lower Effort

The dial's real value isn't "crank it to max to squeeze out points" — it's getting the same score for less. Anthropic's reported SWE-bench Verified figures for Opus 4.5 show this well.

  • medium effort: Opus 4.5 matched Sonnet 4.5's best score — using 76% fewer output tokens.
  • high effort: Opus 4.5 exceeded Sonnet 4.5 by 4.3 percentage points — still using 48% fewer tokens.

Two implications to take away:

  1. Running a smarter model at low effort beats running a weaker model at maximum — and costs less. Model choice and effort choice have to be considered together.
  2. Raising effort is not automatically a token explosion. The same task can finish better and cheaper, because better reasoning cuts wasted, flailing attempts.

So effort is not a simple "burn cost to buy quality" trade — it's an optimization where, depending on the model+effort combination, quality and cost can improve together.

4. Overthinking — Where More Thinking Hurts

Push the dial past the flat region and extra thinking isn't just unhelpful — it can hurt.

  • On narrow-output tasks (classification, extraction, formatting, single-answer), more thinking doesn't improve the answer. Even Anthropic's effort documentation states that on structured-output tasks max "can lead to overthinking."
  • Excess reasoning sometimes second-guesses a correct answer into a wrong one.
  • And all that extra thinking is billed as cost and latency — with no quality gain.

That's why max (Claude) and the top notches of high/xhigh (OpenAI) are reserved seats, not defaults. Claude Code making max apply to the current session only (Part 4) is the same logic — a guardrail against the top notch silently becoming a standing default.

5. Practical Guide — Matching the Dial to Task Difficulty

The whole series' principle, compressed into one table. This is a starting point for matching the dial to difficulty (then fine-tune on your own eval set).

Task type Claude effort OpenAI reasoning_effort
Simple classification / extraction / formatting / short answer low minimal (GPT-5) / none (GPT-5.5) / low (Codex family)
General workflow / ordinary generation medium medium (default)
Complex reasoning / hard coding high (default) high
Long agentic / coding sessions xhigh xhigh (supported models)
Genuine frontier problems (latency/cost-insensitive) max high/xhigh

Operating tips:

  • Set the starting point deliberately. Claude defaults to high, so light tasks need a step down to cut cost/latency. Sonnet 4.6 in particular runs at high if effort is unset, which can raise latency — make medium your explicit default there.
  • High effort needs a generous output cap. At Claude xhigh/max, set a large max_tokens (start ~64k) so the model has room to unfold (Part 3).
  • Output length is not an effort problem. On OpenAI, "think deeply but answer briefly" keeps reasoning_effort and lowers verbosity. On Claude, control length via the prompt.
  • Shallow reasoning → the dial, not the prompt. If a hard task comes back shallow, raising effort a notch is more direct than a "think harder" prompt (Part 3).

6. Loop Cost Is Separate — Task Budgets

Set per-turn depth with effort, but bound the cumulative tokens of a whole agent loop with a separate knob. Claude's Task Budgets (output_config.task_budget, minimum 20,000) tells the model its loop-wide budget and lets it self-moderate against a shrinking countdown — unlike max_tokens, an enforced ceiling the model is unaware of, this is a budget the model knows about and manages (Part 3). On long autonomous work, splitting "depth via effort, total via Task Budget" lets you manage quality and cost at once.

Closing

A reasoning mode is, in the end, the answer to one question — how much compute should this call spend?

  • Principle (Part 1): thinking is test-time compute, billed as output tokens.
  • Claude (Parts 2-4): adaptive thinking sets "whether and when," effort sets "how deep, across the whole response." Claude Code exposes the same dial in the CLI.
  • OpenAI (Part 5): reasoning_effort (depth) and verbosity (length) as two knobs.
  • Choice (Part 6): match the dial to task difficulty. A smarter model at the right effort — that's the path to quality and cost together.

Once you can see the dial, "why is the same model fast and cheap one day, slow and expensive the next" stops being a mystery. It's usually a question of where the dial was set.

Comparisons and figures are grounded in Anthropic and OpenAI primary documentation and Anthropic's reported Opus 4.5 SWE-bench Verified results. Benchmark numbers vary by model and setup — validate on your own eval set before relying on them in production.

Series overview: Series index

๋Œ“๊ธ€

๋Œ“๊ธ€ ์“ฐ๊ธฐ

์ด ๋ธ”๋กœ๊ทธ์˜ ์ธ๊ธฐ ๊ฒŒ์‹œ๋ฌผ

"ML Foundations (9/9) — PyTorch vs TensorFlow, and the Road to Local LLMs"

์ด๋ฏธ์ง€
← 8/9 Deep Learning Architec… ๐Ÿ“š Series Index (series end) The final part. So far we've focused on models. Now we focus on the tools that actually run them . We'll lay out the philosophical difference between PyTorch and TensorFlow, and trace how Hugging Face transformers , llama.cpp , MLX , and Ollama built the bridge to running large language models on your own machine. By the end you should have the full mental model of "download a pretrained LLM and serve it locally." 0. Learning Objectives Explain the eager-vs-graph difference between PyTorch and TensorFlow. Explain, in graph terms, how autograd automates the backward pass. Use Hugging Face transformers ' AutoModel , AutoTokenizer , and pipeline abstractions. Describe llama.cpp's GGUF quantization, INT4 inference, and CPU-first flow. Describe Apple MLX's use of unified memory and how it differs from PyTorch. Run a local LLM with Ollama and call it through an OpenAI-compatible API. ...
Read more »

Agent Memory Engine (2/10) — Building an AI Agent Memory System with SQLite Alone

← 1/10 SQLite Memory Engine f… ๐Ÿ“š Series Index 3/10 memcore Completed → A memory engine that runs without daemons, without dependencies, anywhere ํ•ต์‹ฌ ์š”์•ฝ This post covers how to design an AI agent memory engine (memcore) on a single-file SQLite backend. - Structure: 25 modules, ~6,464 lines, 22 DB tables - Principles: no daemon, single-file portability, no required external dependencies, host-agnostic - Core techniques: 4-layer hybrid search (topic · vector · FTS5 · LIKE), U-tag dialectic-based confidence evolution, ingestion-layer bias rejection rules What This Post Covers One approach to building AI agent memory without a server or cloud vector DB. Specifically: (1) the limitations of text-file-based memory, (2) vector DB alternatives compared and selection criteria, (3) how to layer search tiers, (4) how to evolve confidence scores over time, and (5) how to block biased memory accumulation at the ingestion stage. Design Principles: Why Single-File SQLite Text-file-based ...
Read more »

"ML Foundations (8/9) — Deep Learning Architectures: CNN, RNN, Attention"

์ด๋ฏธ์ง€
← 7/9 Deep Learning Training ๐Ÿ“š Series Index 9/9 PyTorch vs TensorFlow → What did we keep adding on top of an MLP? Vision went CNN, sequences went RNN/LSTM, and eventually both converged on Attention and the Transformer. Each architecture is easier to remember if you read it as what it gave up to gain something else . This part is the one-line summary of the decisive inflection points: LeNet → AlexNet → ResNet → LSTM → Attention → Transformer. 0. Learning Objectives Explain why CNN's convolution, pooling, and weight sharing match images so well. Trace what got unblocked from LeNet → AlexNet → VGG → ResNet as depth grew. State BPTT for RNNs and the vanishing/exploding gradient problem. Write the LSTM gate equations (forget, input, output) with intuition. Write the attention formula in query/key/value form and the scaled dot-product variant. State the precise reasons Transformers replaced RNNs (parallelism + long-range dependencies). 1. ํ•ต์‹ฌ ์š”์•ฝ CNN : weight sharin...
Read more »

"RAG Core Study (14/26) — Evaluation Sets with RAGAS & DeepEval"

์ด๋ฏธ์ง€
Series overview: Series index ๐Ÿ“š Series Index Without an evaluation set, every RAG improvement is a story, not evidence. RAG systems fail in more than one place: retrieval, context selection, answer generation, and grounding. That means "looks better" is not a metric. Part 14 explains how to build a golden dataset , how RAGAS and DeepEval fit into the workflow, and why teams must separate retrieval quality from answer quality if they want tuning decisions to hold up. 0. Prerequisites Part 13 reranking — system quality now depends on multi-stage ranking. Part 1 grounding — a correct-looking answer is not enough. Part 12 Hybrid — multiple retrieval settings need comparison on fixed data. 1. Learning Objectives Build a small but useful golden dataset for RAG. Distinguish retrieval metrics from answer metrics . Use RAGAS and DeepEval in the right roles. Avoid the common traps in judge-model-based evaluation. 2. ํ•ต์‹ฌ ์š”์•ฝ An evaluation set for RAG is ...
Read more »

"ML Foundations (7/9) — Deep Learning Training: Optimizers, Regularization, Initialization"

← 6/9 Neural Networks ๐Ÿ“š Series Index 8/9 Deep Learning Architec… → Building an MLP, as we did in Part 6, is not the same as getting it to train well . The right optimizer, regularization, initialization, and learning rate — when these line up, deep networks converge. When they don't, the network refuses to learn at all. This part is the catalog of those crafts, with formulas, paper citations, and code in one place. 0. Learning Objectives Compare and write the update rules for SGD, Momentum, Nesterov, and Adam. Explain how Dropout, BatchNorm, and LayerNorm work and where they belong in a model. Derive the variance formulas for Xavier and He initialization and match them to activations. Implement step, cosine, and warmup learning-rate schedules in PyTorch. Explain why gradient clipping is effectively required for RNNs and Transformers. Diagnose the most common training failures (NaN, plateau, overfitting) and apply first-line fixes. 1. ํ•ต์‹ฌ ์š”์•ฝ SGD : \(w \leftarro...
Read more »

AI Agents I Built (5/7) — Building an Automated Blogger API Publishing System

← 4/7 My Life Organizer Agent ๐Ÿ“š Series Index 6/7 Dual → Designing a Markdown-Based Auto-Publish Pipeline with Blogger API v3 + OAuth 2.0 Key Summary Blogger API v3 + OAuth 2.0 enables a pipeline that converts Markdown files into publishable HTML posts blogger_publish.py handles the full flow — frontmatter parsing → body cleanup → HTML conversion → CSS injection → API post — in a single script For bulk publishing, quota management (delay, retry, TOC suppression) is the critical design concern Platform Selection Criteria API access is mandatory for AI agent-driven content publishing. A comparison of platforms: Platform API Cost AdSense Decision WordPress.com REST API Paid plan required Paid only Cost overhead Ghost REST API Self-hosted or paid Manual setup Ops overhead Blogger v3 API Free Native integration Adopted Blogger is free, exposes a REST API, and integrates natively with AdSense. For personal blog automation, this combination i...
Read more »

"ML Foundations (6/9) — Neural Networks: From Perceptron to MLP"

์ด๋ฏธ์ง€
← 5/9 Regularization & Model… ๐Ÿ“š Series Index 7/9 Deep Learning Training → A neural network is logistic regression from Part 4 stacked into layers. Each layer passes through a new nonlinear activation, expanding the family of functions the network can represent. In this part we put how a neural network learns onto one page. Perceptrons, XOR, MLPs, backpropagation, activation functions, and the universal approximation theorem all meet here. 0. Learning Objectives Write the perceptron's learning rule and explain why it cannot solve XOR. Express an MLP's forward pass as one line of matrix products. Derive backpropagation by hand on a two-layer MLP. Compare the definitions, derivatives, and trade-offs of sigmoid, tanh, ReLU, and GELU. State the universal approximation theorem (Cybenko 1989 / Hornik 1991) precisely. Train an MLP end-to-end in PyTorch. 1. ํ•ต์‹ฌ ์š”์•ฝ Perceptron (1958): \(\hat{y} = \mathrm{step}(w^\top x + b)\). Converges only on linearly separabl...
Read more »