Anthropic Releases Official Cost-Optimization Guide for Claude Code

Anthropic Releases Official Cost-Optimization Guide for Claude Code

Just now, Anthropic published an official blog post with a clear and urgent message: “Stop wasting tokens — we can’t watch you overspend anymore.”

The post reveals that many developers are unintentionally inflating their Claude Code usage costs by orders of magnitude — not due to model limitations, but through suboptimal interaction patterns. To help users maximize value, Anthropic details six actionable, engineering-first optimization strategies.

✅ Six Proven Cost-Saving Techniques

1. /clear after every task
   Reset context immediately after resolving a bug or completing a task. Avoid carrying forward irrelevant files, command outputs, or exploration history into the next session — each retained token incurs cost in every subsequent round-trip.

2. Lock model and effort level upfront
   Switching models (e.g., Sonnet → Opus) or adjusting inference intensity (/effort) invalidates the entire prompt cache. Prefill computations must be re-run at full price — no discount applies.

3. Use @file references — never raw paths
   Attach files directly via @filename.txt. This bypasses costly tool-calling overhead and prevents Claude from blindly scanning multiple files (which bloats history with failed attempts).

4. Enable quiet flags for verbose commands
   Configure CLI tools like Vitest with compact reporters (e.g., --reporter=dot). Replace 400-line test logs with 3-line summaries — dramatically reducing output token volume.

5. Run /compact before idle breaks
   Compressing active cache costs just 10% of normal prefill. Wait until cache expires? You’ll pay full price to reload and compress — a 10× penalty.

6. Delegate heavy-output tasks to subagents
   Subagents operate in isolated contexts. They execute file reads, command runs, and analysis internally — only returning concise conclusions. No intermediate artifacts pollute your main conversation.

💡 Understanding Token Economics

Every Claude Code interaction involves two distinct phases:

• Prefill (input tokens): Parallel GPU processing of system prompts, CLAUDE.md, messages, and full conversation history.
• Decode (output tokens): Serial token-by-token generation — including reasoning steps, tool calls, and final output.

⚠️ Critical insight: Output tokens cost up to 5× more than input tokens — because decoding is computationally serial and latency-sensitive.

Model	Input ($/M tokens)	Output ($/M tokens)
Opus 5	$5.00	$25.00
Sonnet 5	$2.00	$10.00
Haiku 4.5	$1.00	$5.00

💡 Rule of thumb: Use Sonnet for routine tasks; reserve Opus only for high-stakes, complex reasoning.

🧠 Prompt Caching: Your Biggest Leverage Point

Claude Code caches identical request prefixes — saving 90% on prefill cost (just 0.1× base rate). But cache hits require byte-perfect prefix matches. Six common pitfalls break it:

• 🔁 Switching models (/model opus)
• ⚙️ Changing effort level (/effort max)
• ⚡ Toggling Fast Mode
• 📄 Running /compact mid-session
• ⏳ Cache expiration (1 hr for subscribers, 5 min for API users)
• 🔄 Resuming stale sessions

✅ Pro tip: Lock model + effort at session start, defer /compact to idle time, and avoid opusplan mode’s repeated model switching.

📉 Why Your Session Is Silently Getting Heavier

Each file read or command output appends content to your history. By round #40, you’re re-sending all prior rounds’ cumulative tokens — resulting in near O(n²) growth.

Claude Code caps raw output at 30,000 chars (then uses temp files), but smaller outputs — like 400-line test reports — persist fully in history. Every round carries them forward.

Smart “Diet” Tactics:

• ✅ @file instead of manual path lookup
• ✅ Quiet CLI flags (--reporter=dot)
• ✅ Subagents for log scanning / large-file review
• ✅ /clear between unrelated tasks
• ✅ /rewind to discard missteps without breaking cache

🌐 Token Awareness: A New Developer Competency

This isn’t just about saving dollars — it’s about cultivating LLM-native engineering discipline. Anthropic’s own engineers report:

• 80% of production code written with AI
• 8× increase in merged PRs/year
• 52× acceleration in benchmark iteration

Without token hygiene, those gains would vanish under runaway inference costs.

🎯 The core skill isn’t using AI — it’s orchestrating it: selecting models wisely, guarding cache integrity, managing context weight, and calibrating inference depth. That’s how $3 tasks stay $3 — not $30.

---

Source: Anthropic Blog: Maximizing the Value of Your Claude Code Sessions