"LLM Reasoning Modes (2/6) — Claude's Thinking: From Fixed Budget to Adaptive"

Part 1 established that thinking costs. Part 2 looks at how Claude controls that thinking. The pivot is the move from a fixed token budget (budget_tokens) to adaptive thinking, where the model decides for itself.

Claude's reasoning control has two layers: Thinking (whether to think, and how deep) and effort (the token budget for the whole response). They work together but are separate knobs. Part 2 covers the first — Thinking — and effort is Part 3.

In One Paragraph

Old Claude gave thinking a fixed token budget via thinking: {type: "enabled", budget_tokens: N}. Current models dropped that for adaptive thinking (thinking: {type: "adaptive"}) — the model decides when and how much to think per request. budget_tokens is deprecated (still works) on Opus 4.6 / Sonnet 4.6, and removed (400 error) on Opus 4.7, 4.8, and Fable 5. Thinking depth is now controlled by effort. The raw chain of thought is never returned; display only toggles a summary.

1. Extended Thinking: The Fixed-Budget Era

Claude's first way of exposing thinking was an explicit token budget.

response = client.messages.create(
    model="claude-sonnet-4-5",
    max_tokens=16000,
    thinking={"type": "enabled", "budget_tokens": 8000},
    messages=[...],
)

The rules were simple.

  • budget_tokens must be less than max_tokens.
  • Minimum 1024.
  • The developer pins "this call may spend up to N tokens on thinking."

The problem is that this is rigid. Reserving a budget on an easy question wastes it; too small a budget on a hard question truncates the thinking. The developer had to estimate difficulty and set a number per call.

2. Why the Fixed Budget Was Dropped — Adaptive Thinking

Current Claude uses adaptive thinking.

response = client.messages.create(
    model="claude-opus-4-8",
    max_tokens=16000,
    thinking={"type": "adaptive"},
    output_config={"effort": "high"},
    messages=[...],
)

The key difference is who decides. A fixed budget lets the developer pin thinking volume; adaptive lets the model look at the request and decide when and how much to think. It thinks briefly or skips on easy questions and thinks longer on hard ones.

There's a side effect too. Adaptive thinking automatically enables interleaved thinking (thinking between tool calls) — no beta header required. In an agent loop, the "look at the tool result, plan the next step" reasoning slots in naturally.

So what controls thinking depth? Effort (Part 3). Adaptive thinking leaves "whether and when" to the model; effort sets the tone for "how deep." Their combination replaces the single fixed budget.

3. Per-Model Status — What Works and What's Blocked

This transition is staged differently per model. Here's the table you'll hit most often when migrating.

Model budget_tokens (fixed budget) Adaptive thinking Notes
Opus 4.5 Used (manual thinking) — effort works alongside the thinking budget
Opus 4.6 Deprecated (still allowed) Recommended transitional escape hatch only
Sonnet 4.6 Deprecated (still allowed) Recommended defaults to high effort
Opus 4.7 Removed → 400 Used thinking is off if not specified
Opus 4.8 Removed → 400 Used same request surface as 4.7
Fable 5 Removed → 400 Always on omit thinking entirely; disabled also 400s

How to read it:

  • Opus 4.7 / 4.8: you must explicitly set thinking: {type: "adaptive"} for thinking to turn on. Omitting the field runs without thinking. Sending budget_tokens returns 400.
  • Fable 5: thinking is always on. Omit the thinking field entirely; {type: "disabled"} returns 400.
  • Opus 4.5: the one model that uses manual thinking (budget_tokens) together with effort.

For new code the answer is singular — don't use budget_tokens; go with thinking: {type: "adaptive"} + effort. The very concept of a "fixed thinking budget" is on its way out.

4. Thinking Blocks and display — Visible vs Hidden

Thinking happens and is billed, but the raw text is never returned. The response stream contains a thinking-type block, and what it carries is decided by thinking.display.

  • display: "summarized" → a readable summary is included.
  • display: "omitted" → the thinking block arrives but its text is an empty string.

The trap: the default is omitted (on Fable 5, Mythos 5, Opus 4.8, and 4.7). This is a silent change from Opus 4.6, where it was summarized. As a result, a UI that meant to show progress can suddenly look like "a long pause, then the answer." To surface a summary, set it explicitly.

thinking={"type": "adaptive", "display": "summarized"}

display controls visibility only. Under any setting, thinking happens and is billed the same. And under no setting is the raw chain of thought exposed — a summary is the ceiling.

5. Multi-Turn Rule — Echo Thinking Blocks Back Unchanged

In agents and multi-turn there's one rule to keep. When continuing on the same model, send the thinking blocks back exactly as received, unmodified. Even an empty-text block goes back as-is — the API rejects modified blocks, not blocks you've read.

Conversely, when handing off to a different model, those thinking blocks are dropped from the prompt (usually silently, not an error). The drop happens before billing, so it costs nothing and there's nothing to strip out. Ordinary thinking blocks replay freely across models; treat this drop as normal behavior.

6. Thinking and effort (Next Part Preview)

To summarize, current Claude's reasoning control reads like this.

  • adaptive thinking = the model decides "whether and when to think."
  • effort = the tone for "how deep to think, and how many tokens to spend across the whole response."

At high, xhigh, and max effort the model almost always thinks deeply; at lower effort it may skip thinking on easy problems. So effort dials adaptive thinking's eagerness up or down.

Part 3 takes effort head-on — what low / medium / high / xhigh / max each change, not only in thinking but in tool calls and preamble.

Parameter behavior and per-model support are grounded in Anthropic's primary documentation (extended thinking, adaptive thinking, effort) and the model migration guide.

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 »