"Claude Code Token & Cache Cost Breakdown 2026 — The 1.25× vs 2× Multiplier Trap"

Official pricing, cache minimums, /usage tracking, and 10 ways to cut tokens

핵심 요약

  • Audience: Claude Code users who were surprised by a monthly bill, or teams estimating budget before rollout.
  • What you'll get: Current (April 2026) per-model pricing, cache multipliers (1.25× / 2× / 0.1×), per-model minimum cacheable tokens, /usage for live tracking, the official 10-item reduction playbook, and the hidden costs (extended thinking, agent teams, background).
  • Prerequisite: Claude Code installed (install guide). Pro/Max subscribers pay a flat rate — the numbers below don't apply to your bill, only to usage-bar planning.

1. Two billing axes — tokens and cache

Claude Code charges based on Anthropic API token consumption. Cost comes from two axes.

  1. Input / output tokens — every turn bills separately for the prompt (input) and the response (output).
  2. Prompt cache — caching the system prompt, CLAUDE.md, or long context makes retransmissions of the same content cost 0.1×. But cache writes are more expensive than standard input (1.25× or 2×).

Without grasping these two axes, you can't explain "why did the second response cost less than the first?"

Pro/Max/Team/Enterprise subscribers pay a flat monthly rate — per-token costs don't map to billing. What matters is the usage-bar limit per plan. The table below is for Anthropic Console pay-as-you-go.

2. Official pricing (April 2026, per million tokens)

Per Anthropic's pricing page.

Model Input 5-min write 1-hour write Cache read Output
Opus 4.7 $5 $6.25 $10 $0.50 $25
Opus 4.6 $5 $6.25 $10 $0.50 $25
Opus 4.5 $5 $6.25 $10 $0.50 $25
Sonnet 4.6 $3 $3.75 $6 $0.30 $15
Sonnet 4.5 $3 $3.75 $6 $0.30 $15
Haiku 4.5 $1 $1.25 $2 $0.10 $5
Haiku 3.5 $0.80 $1.00 $1.60 $0.08 $4

Observations: - Output is 5× input. Long responses drive cost fastest. - Cache writes are a one-time premium. Reads that follow are 0.1×. - Opus is 5× Sonnet. Model choice is a major lever.

3. Cache multiplier structure — 1.25× vs 2× vs 0.1×

Official rules:

Operation Multiplier Meaning
5-min TTL write (default ephemeral) 1.25× input Wins if reused within 5 minutes
1-hour TTL write (extended) 2× input Wins for 5-min to 1-hour gaps
Cache read (hit/refresh) 0.1× input 90% discount on retransmitting same content

Example — Sonnet 4.6, a 50,000-token context reused 10 times in one session:

  • No cache: 50,000 × 10 × $3/MTok = $1.50 (input only)
  • 5-min cache: first write 50,000 × $3.75/MTok + reads 50,000 × 9 × $0.30/MTok = $0.19 + $0.135 = $0.325 (~78% savings)
  • 1-hour cache: first write 50,000 × $6/MTok + reads 50,000 × 9 × $0.30/MTok = $0.30 + $0.135 = $0.435 (~71% savings)

If you reuse within 5 minutes, 5-min TTL always wins. But idle over 5 minutes invalidates the cache and you pay full write cost again — 1-hour wins in that regime.

3.1 What TTL does Claude Code use?

Caveat: The official /costs page doesn't explicitly specify Claude Code's default TTL. Observationally, Claude Code targets cross-session resume use, behaving as a 1-hour TTL-only client. Budget with the 2× multiplier to be safe.

So Claude Code's upside is "cache hits survive long sessions". The downside is "cache write premium is 2×, not 1.25×".

4. Per-model minimum cacheable tokens

If the cacheable portion of a request is below the minimum, caching silently doesn't apply (no error, just bypassed).

Model Minimum cacheable tokens
Opus 4.7 / 4.6 / 4.5 4,096
Sonnet 4.6 2,048
Sonnet 4.5 1,024
Haiku 4.5 4,096
Haiku 3.5 2,048

Using Opus on a project with a short CLAUDE.md often misses the 4K threshold. In that regime Sonnet is more economical.

4.1 Cache breakpoint limit

Max 4 cache_control breakpoints per request. You decide where to draw cache boundaries (system prompt / tool definitions / long files). As a user you don't configure this directly — Claude Code handles it internally.

5. Track live cost with /usage

Run /usage (aliases /cost, /stats) in-session to see tokens, cost, and code change counts.

Official example:

Total cost:            $0.55
Total duration (API):  6m 19.7s
Total duration (wall): 6h 33m 10.2s
Total code changes:    0 lines added, 0 lines removed

The dollar figure is locally estimated from token counts. Authoritative billing is on the Anthropic Console Usage page.

5.1 Show on status line

/statusline can keep context/tokens/cost in the terminal bar continuously — recommended.

/statusline

Interactive setup; can auto-configure from your shell prompt.

5.2 Enterprise averages (reference)

Published numbers: - Average ~$13/developer/active day - Monthly $150–250/developer - 90% of users stay under $30/active day

These are enterprise API deployments. Individual Pro/Max experience varies significantly.

6. The 10-item official reduction playbook

6.1 /clear between tasks

When switching to unrelated work, /clear. Stale context wastes tokens on every subsequent message.

6.2 /compact with focus instructions

/compact Focus on code samples and API usage

Telling Claude what to preserve avoids losing key context.

Or set a default in CLAUDE.md:

When you are using compact, please focus on test output and code changes

6.3 Model choice — Sonnet default, Opus when complex

Sonnet handles most coding well and costs 5× less than Opus. Reserve Opus for architectural decisions or multi-step reasoning.

6.4 Cut MCP server overhead

MCP tool definitions are deferred by default — only tool names enter context until Claude invokes a specific tool. Use /context to see what's consuming space. Disable unused servers with /mcp.

Prefer CLI tools (gh, aws, gcloud, sentry-cli) — they add no per-tool listing.

6.5 Code intelligence plugins

For typed languages, install a code intelligence plugin. A "go to definition" call replaces grep + multiple file reads.

6.6 Preprocess with hooks

A hook can grep ERROR from a 10,000-line log before Claude ever sees it. Official example — PreToolUse hook filtering test output to failures only:

.claude/settings.json:

{
  "hooks": {
    "PreToolUse": [
      {
        "matcher": "Bash",
        "hooks": [{"type": "command", "command": "~/.claude/hooks/filter-test-output.sh"}]
      }
    ]
  }
}

filter-test-output.sh:

#!/bin/bash
input=$(cat)
cmd=$(echo "$input" | jq -r '.tool_input.command')

if [[ "$cmd" =~ ^(npm test|pytest|go test) ]]; then
  filtered_cmd="$cmd 2>&1 | grep -A 5 -E '(FAIL|ERROR|error:)' | head -100"
  echo "{\"hookSpecificOutput\":{\"hookEventName\":\"PreToolUse\",\"permissionDecision\":\"allow\",\"updatedInput\":{\"command\":\"$filtered_cmd\"}}}"
else
  echo "{}"
fi

Tens of thousands of tokens become hundreds. Full hook mechanics: hooks complete guide.

6.7 Move CLAUDE.md into skills

CLAUDE.md is loaded every session. Move detailed workflow instructions (PR reviews, DB migrations) to skills — they load on demand. Keep CLAUDE.md under 200 lines.

6.8 Tune extended thinking

Extended thinking is on by default. Thinking tokens bill as output tokens — default budget can be tens of thousands per request. For simple tasks, this is waste.

/effort low
MAX_THINKING_TOKENS=8000 claude

6.9 Delegate verbose work to subagents

Test runs, documentation fetches, and log parsing consume significant context. Delegate these to subagents — verbose output stays in the subagent's context, only a summary returns.

6.10 Write specific prompts

"Improve this codebase" triggers broad scanning. "Add input validation to the login function in auth.ts" finishes with minimal file reads.

7. Three hidden costs

7.1 Background

Claude Code uses some tokens while idle: - Conversation summarization jobs (for claude --resume) - Status-checking commands

Official: under $0.04 per session. Small, but multiplies across many open sessions.

7.2 Agent teams

A feature-flag capability (CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1) that spawns multiple instances. Each teammate runs its own context window, so plan-mode costs are ~7× a standard session.

Mitigation: - Use Sonnet for teammates (never Opus) - Keep teams small - Keep spawn prompts focused - Clean up when done

7.3 Default extended thinking on Opus

Complex work on Opus 4.7 can push thinking to hundreds of thousands of tokens. Make checking /usage a habit.

8. Team rollout — TPM/RPM recommendations

Official per-user recommendations by org size (Anthropic API plans):

Team size TPM/user RPM/user
1-5 200k-300k 5-7
5-20 100k-150k 2.5-3.5
20-50 50k-75k 1.25-1.75
50-100 25k-35k 0.62-0.87
100-500 15k-20k 0.37-0.47
500+ 10k-15k 0.25-0.35

Larger teams get lower per-user allocations because concurrency drops. Limits apply at the org level, so individuals can temporarily exceed their share.

9. PostToolUse hook for cost logging

/usage shows the current session only. For long-term tracking, append to a file with a hook.

.claude/hooks/log-cost.sh:

#!/bin/bash

INPUT=$(cat)
TOOL_NAME=$(echo "$INPUT" | jq -r '.tool_name')
SESSION_ID=$(echo "$INPUT" | jq -r '.session_id')
TS=$(date +%s)

echo "$TS | $SESSION_ID | $TOOL_NAME" >> "$CLAUDE_PROJECT_DIR/tasks/log/tool-usage.tsv"

exit 0

.claude/settings.local.json:

{
  "hooks": {
    "PostToolUse": [
      {
        "matcher": "*",
        "hooks": [
          {"type": "command", "command": "\"$CLAUDE_PROJECT_DIR\"/.claude/hooks/log-cost.sh"}
        ]
      }
    ]
  }
}

The example only counts tool calls, but parsing /usage output in a PostToolBatch or Stop hook yields precise monthly stats.

10. Counter-scenarios — when this optimization doesn't apply

  • Individual Pro/Max subscribers → flat-rate billing; per-token cost is irrelevant. Only the usage-bar matters. Pro hits a monthly cap → wait/upgrade.
  • Short sessions only → below the cache minimum (Opus 4K, Sonnet 2K), caching doesn't kick in. No effect from optimization.
  • Prototyping / exploration → speed beats cost. Overusing /compact or model switching disrupts thinking flow.
  • Enterprise API deployment → team-wide tracking is better handled by a gateway (LiteLLM per official docs) than per-user hooks.

11. What's next

With cost structure in hand:

  1. Hooks complete guide — the foundation for preprocessing and cost-logging hooks above.
  2. CLAUDE.md authoring guide — keep it under 200 lines; move details to skills.
  3. Slash commands complete reference — /usage, /context, /compact, /clear, /model, /effort all documented.

References

This is post 5/15 in the "AI Coding CLI Entry Guide" series. last verified: 2026-04-25 (per Anthropic official pricing and Claude Code Costs docs).

댓글

댓글 쓰기

이 블로그의 인기 게시물

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

← 1/10 SQLite Memory Engine f… 📚 Series Index 3/10 memcore Completed → A memory engine that runs without daemons, without dependencies, anywhere 핵심 요약 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 (9/9) — PyTorch vs TensorFlow, and the Road to Local LLMs"

이미지
← 8/9 Deep Learning Architec… 📚 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 »

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

이미지
Series overview: Series index 📚 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. 핵심 요약 An evaluation set for RAG is ...
Read more »

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

이미지
← 7/9 Deep Learning Training 📚 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. 핵심 요약 CNN : weight sharin...
Read more »

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

← 6/9 Neural Networks 📚 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. 핵심 요약 SGD : \(w \leftarro...
Read more »

OpenClaw to Hermes Migration (2/13) — What to Preserve, Partially Port, or Discard

← 1/13 Current Structure Inve… 📚 Series Index 3/13 What Moves to Hermes → A code review record classifying operational scripts into three tiers: preserve, partial-preserve, and discard. What This Post Covers An asset classification framework (preserve / partial-preserve / discard) for migrating a legacy agent system to a successor A concrete case applying the Strangler Fig + Subset Migration strategy to a memory pipeline Classification results evaluated across LOC, external dependencies, and replaceability axes A method for slimming down a bloated single file (1,310 LOC) using the migration as a natural entry point Problem Definition: Port Everything or Cut Deep? The OpenClaw inventory comprises multiple agents, multiple scheduler/daemon processes, and multiple operational scripts. When moving to the successor system, Hermes, the choice is not binary. A full port carries accumulated technical debt along with it; a minimal port discards hard-won domain logic. Code-revie...
Read more »

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

← 4/7 My Life Organizer Agent 📚 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 »