惯性聚合 高效追踪和阅读你感兴趣的博客、新闻、科技资讯
阅读原文 在惯性聚合中打开

推荐订阅源

S
SegmentFault 最新的问题
博客园 - 三生石上(FineUI控件)
WordPress大学
WordPress大学
博客园 - 【当耐特】
月光博客
月光博客
Vercel News
Vercel News
D
Docker
I
InfoQ
Apple Machine Learning Research
Apple Machine Learning Research
博客园 - 叶小钗
MongoDB | Blog
MongoDB | Blog
GbyAI
GbyAI
有赞技术团队
有赞技术团队
雷峰网
雷峰网
博客园 - 聂微东
小众软件
小众软件
Y
Y Combinator Blog
腾讯CDC
L
LangChain Blog
The GitHub Blog
The GitHub Blog
宝玉的分享
宝玉的分享
Stack Overflow Blog
Stack Overflow Blog
大猫的无限游戏
大猫的无限游戏
T
The Blog of Author Tim Ferriss

DEV Community

Authentication Security Deep Dive: From Brute Force to Salted Hashing (With Java Examples) Why AI Systems Don’t Fail — They Drift Spilling beans for how i learn for exam😁"Reinforcement Learning Cheat Sheet" I Replaced Chrome with Safari for AI Browser Automation. Here's What Broke (and What Finally Worked) How Python Borrows Other People's Work The $40 Architecture: Processing 1 Billion API Requests with 99.99% Uptime Vibe Coding: A Workflow Guide (From Zero to SaaS) Most webhook security guides protect the wrong side. The scary part is delivery. Headless CMS for TanStack Start: Build a Blog with Cosmic EU Age Verification App "Hacked in 2 Minutes" — What Actually Happened Comfy Cloud’s delete function does not actually remove files Running AI Models on GPU Cloud Servers: A Beginner Guide Event-driven media intelligence with AWS Step Functions and Bedrock I scored 500 AI prompts across 8 quality dimensions — here's what broke How to Call Google Gemini API from Next.js (Free Tier, No Backend Needed) The Portal Protocol: Reclaiming Human Connection in the Age of AI How to Fix Your Team's Scattered Knowledge Problem With a Self-Hosted Forum Intro to tc Cloud Functors: A Graph-First Mental Model for the Modern Cloud Designing Multi-Tenant Backends With Both Ownership and Team Access I Built a Neumorphic CSS Library with 77+ Components — Here's What I Learned PostgreSQL Performance Optimization: Why Connection Pooling Is Critical at Scale Cómo construí un SaaS multi-rubro para gestionar expensas en Argentina con FastAPI + Vue 3 🚀 I Built an Ethical Hacking Scanner Tool – Open Source Project I Replaced /usage and /context in Claude Code With a Single Statusline A Pythonic Way to Handle Emails (IMAP/SMTP) with Auto-Discovery and AI-Ready Design I Collected 8.9 Million Polymarket Price Points — Here's What I Found About How Markets Really Move EcoTrack AI — Carbon Footprint Tracker & Dashboard Everyone's Using AI. No One Agrees How. 5 self-hosted ebook managers worth trying in 2026 Building Your First AI Agent with LangChain: From Chatbot to Autonomous Assistant
I Audited 3 Months of Claude Code Billing — Most Communit...
Recca Tsai · 2026-04-26 · via DEV Community

Recca Tsai

I Audited 3 Months of Claude Code Billing — Most Community Cost-Saving Tips Don''t Work

Originally published at recca0120.github.io

This past week, chasing a vague "quota burns faster lately" feeling, I scanned three months of my own Claude Code logs. ~\$127K equivalent cost, 127K turns, four models, hundreds of sessions.

The uncomfortable finding: the cost-saving tips floating around on Reddit / HN / Twitter mostly don't survive real data. "Sessions are too long, run /clear," "too many skills, prune them," "MCP servers should be lean" — these all sound right. But against three months of actual data, almost none holds up. Only two things actually shrink the bill, and neither is about "optimizing your habits."

Earlier I wrote }}">Scanned 95 days of Claude Code logs, found a second cache TTL silent regression and }}">17-day follow-up covering server-side behavior. This post is the extension: with server behavior confirmed unfixable, what's left for the user side.

Three Months of Bills

A single primary development project (one codebase, solo dev), monthly Claude Code totals:

Month Equiv \$ Dominant Model Key Event
2026-02 \$1,015 Five models mixed Trial period, low volume
2026-03 \$48,623 99.6% Opus 4.6 Heavy usage starts; per-call prefix jumped 58K → 417K in one step
2026-04 \$77,754 Opus 4.6 \$51K + Opus 4.7 \$25K Opus 4.7 release on 4/16, alias auto-upgraded

Two key observations:

  1. From March to April, Opus 4.6 cost barely changed (\$48K → \$51K, +7%). \$/turn went from \$0.692 → \$0.713, a 3% delta. Usage habits stayed flat.
  2. The extra \$25K in April is almost entirely the Opus 4.7 layer.

So the "lately it got expensive" feeling isn't because I changed anything — it's because Opus 4.7 shipped on 4/16 and the opus alias automatically pointed to the new version. With no version pinned in settings, the next session jumped to it.

This is normal alias behavior, not something hidden. But for subscription users, the quota impact is real — as we'll see, the new version's adaptive thinking burns quota at 2.4× the old.

Multi-Dimensional Breakdown for April

Here's the full cut by model for one month:

Dimension Opus 4.6 Opus 4.7 Sonnet 4.6 Haiku
Volume
Sessions (main/sub) 24/138 18/84 5/46 1/376
Total turns 72,431 31,621 15,182 16,138
% of total turns 47.4% 20.7% 9.9% 10.6%
Wall-clock hours 635 270 72 14
Active hours (no idle) 237.9 106.5 40.2 6.2
Output Profile
Turns/active hour 305 297 378 2,614
Tools/turn 0.62 0.63 0.64 0.68
Output tokens/turn 227 667 456 101
Sub:Main turn ratio 1:1.32 1:15.56 1:14.95 n/a
Cost
Equivalent \$ \$51,700 \$24,595 \$773 \$114
Cost share 67.0% 31.9% 1.0% 0.1%
Quota burn rate 1.0× 2.4× 0.2× 0.05×
\$/turn \$0.714 \$0.778 \$0.051 \$0.007

Cross-column observations:

Opus 4.7 emits 2.9× more output tokens per turn (667 vs 227). It's not verbose — adaptive thinking's reasoning chain counts as output. To complete the same task, 4.7 burns roughly 3× the output of 4.6.

Opus 4.7 doesn't delegate. Sub:Main turn ratio jumped from 4.6's 1:1.32 to 1:15.56 — 4.6 is a "give half to sub-agents" collaborator, 4.7 is a "think it through alone" lone wolf. This explains the 3× output per turn: thinking is all done in-house.

Sonnet 4.6 \$/turn is 1/16 of Opus. But it only made up 9.9% of turns — clearly underused.

Haiku is the invisible workhorse. Zero main sessions, 376 sub-sessions, 16K turns for \$114 — all triggered automatically by Claude Code's built-in Explore / Plan agents. Untouched, still doing 10% of total turns.

Five Common "Cost-Saving Tips" Debunked

Community lore (Reddit / HN / Discord) graded against real data.

❌ "Long sessions are the culprit"

The intuition: longer sessions mean longer conversation history, more cache prefix re-read per turn, more cost as the session drags on.

The data: March vs April Opus 4.6 usage is nearly identical (69,980 vs 72,510 turns), but \$/turn moved from \$0.692 → \$0.713, a 3% bump. If long sessions were the driver, the per-turn cost should creep up month over month. It doesn't.

More precisely: cache_read accounts for 77–88% of cost on both Opus versions. The number is huge, but the ratio has been that way since heavy Claude Code usage started — it's the inherent cost of "talking to an LLM," not the price of "not splitting sessions." /clear doesn't recover much.

❌ "Run /clear after 5+ min idle"

The intuition: 5-minute cache TTL means a brief idle expires the cache, so the next turn pays for a rewrite.

The data: my }}">second audit shows the main agent has been writing 100% to 1h TTL for 17 straight days since 4/9, with zero 5m writes. Idle a while and come back, cache is still there. No extra write cost.

The forced 5m downgrade only hits sub-agents (same post). But sub-agents only contributed a small slice of April's cost (~\$1,500 estimated), two orders of magnitude less than the \$25K from 4.7.

❌ "Too many skills"

The intuition: loaded skills inject metadata into the system prompt every turn.

The data: I actually measured. 40 skill descriptions add up to ~5–10K tokens. In a 425K per-call prefix, that's 1–2%. Deleting all of them saves <\$1K/month — not worth the effort.

❌ "Too many MCP servers"

The intuition: MCP tool definitions land in the prefix every turn.

The data: setup is 3–4 MCPs (pixel-mcp, the Google Workspace trio), several of which fail to connect and don't load. Already lean, nothing to trim.

❌ "CLAUDE.md is too long"

The intuition: CLAUDE.md gets re-read every turn.

The data: the project root CLAUDE.md is 1 byte (essentially empty), the global one is 0 bytes. Zero impact.

These five aren't wrong in every scenario. For someone with a 50K-token CLAUDE.md or 20 loaded MCP servers, they apply. But as generic advice spread to everyone, data shows they barely help a heavy single-project workflow.

✅ The Two Things That Actually Work

After the intuition reckoning, only two things hold up against the data:

1. Pin Specific Model Versions in settings.json

Don't use opus / sonnet aliases. When Anthropic ships a new version, the alias auto-points to it — invisible to the user but quota behavior shifts dramatically.

{
  "model": "claude-opus-4-6",
  "permissions": { "...your existing..." }
}

Enter fullscreen mode Exit fullscreen mode

This way when opus-4.8 / 4.9 ships, you don't auto-follow. New versions aren't always more economical — for 4.6 vs 4.7:

  • \$/turn +9%
  • Output/turn +190%
  • Quota burn +140%
  • Turns to complete same work only −12%

Net CP value is 1.9× higher on 4.6. Every model release, check cnighswonger's advisory and run your own data for a while before deciding to upgrade.

On adaptive thinking: 4.7 burns hard mainly because adaptive thinking counts the reasoning chain as output tokens. Opus 4.6 / Sonnet 4.6 let you disable it via CLAUDE_CODE_DISABLE_ADAPTIVE_THINKING=1, but Opus 4.7 forces it on with no toggle — that's why "lock 4.6" is the practical fix instead of trying to mitigate 4.7. Auto mode and adaptive thinking are independent features; pinning 4.6 doesn't affect auto mode.

2. Route Review / Fix / Test to Sonnet

The \$/turn gap is real (Opus \$0.71 vs Sonnet \$0.045 — 16×). My April: 14K turns on Sonnet for \$643, same turns on Opus 4.6 would have been \$10K.

Switch to Sonnet for:

  • Code review, reading PR diffs
  • Small bug fixes, type annotations, null checks
  • Writing tests, adding test cases
  • Docs, commit messages, changelogs
  • Renames, simple refactors

Stay on Opus for:

  • Cross-file architectural rewrites
  • Design decisions needing long reasoning chains
  • Complex debugging (race conditions, memory leaks)
  • Exploring an unfamiliar codebase the first time

How: inside a session, /model claude-sonnet-4-6 to switch over for a few rounds, then /model claude-opus-4-6 to switch back. Don't lock Sonnet in settings — you'll forget to switch when you need Opus.

Real Magnitudes

If both levers are in place, expected April-baseline change (against \$77K):

Action Expected Savings % of Monthly Bill
Pin 4.6 (cancel 4.7 auto-follow) \$25K/mo 32%
Route review/fix/test to Sonnet (expand to 30% of turns) \$10–15K/mo 13–20%
Total \$35–40K/mo 45–52%

The remaining 50% is the inherent cost of "heavy Opus 4.6 usage on a primary project" — not optimizable, and shouldn't be. That's the work itself.

Lessons

The biggest takeaway from turning myself into a dataset isn't the savings — it's seeing how unreliable community intuition is.

"Shorter session = cheaper," "fewer skills = cleaner" might hold in some scenarios, but for single-project heavy-use workflows they're flat wrong. Without breaking cost down to model × session × turn, I'd never have spotted that "the Opus 4.7 alias upgrade" is the single biggest reason April got expensive.

Broader lessons:

  1. Floating optimization tips are noise — without data, "cost-saving advice" often optimizes the wrong thing
  2. Aliases hand cost control to the vendor — the mechanism isn't bad, but it's a real risk for subscription users with quota planning
  3. Multi-model strategy beats single-model tuning — same dollar, Sonnet does 16× the turn volume

If you want to scan your own, the 60-line Python from }}">the first post is reusable — adjust the cost calc and you'll get this analysis for your data. Make yourself a dataset and re-check what the community thinks it knows.

}}">Background: Claude Code session cost & cache misconception

References