mana

name	mana
description	Analyze mana (Anthropic rate limit quota) consumption patterns. Use when investigating "where did the mana go", diagnosing cache busts, estimating cost of planned work, or understanding API spend efficiency.

Mana Analysis

Mana = Anthropic's rate limit quota. Resets to 100% every 5 hours. It represents remaining API quota before rate limiting.

Key Concepts

Anthropic plans have a usage limit per 5-hour window. The cost_usd values logged per API call are calculated from Anthropic's per-token list pricing — what the same usage would cost on the pay-per-use API. They're useful for relative comparisons between turns but aren't actual spend on subscription plans.

Mana is the real constraint. It's the rate limit that resets every 5 hours. Each percentage point of mana corresponds to a fraction of your plan's 5-hour budget.

Mana is shared across ALL consumers on the same Anthropic account: foci API, Claude Code, mobile app, desktop app. If mana drops faster than expected, check whether other apps were also consuming.

Quick Diagnostics

"Where did the mana go?"

# Total cost in last N hours
sqlite3 ~/data/api.db "SELECT SUM(cost_usd), COUNT(*) FROM api_calls WHERE ts > datetime('now', '-3 hours')"

# Biggest individual calls
sqlite3 ~/data/api.db "SELECT ts, call_type, cost_usd, cache_read, cache_write FROM api_calls WHERE ts > datetime('now', '-3 hours') ORDER BY cost_usd DESC LIMIT 10"

# Cache busts (cache_read = 0 with large cache_write)
sqlite3 ~/data/api.db "SELECT ts, cost_usd, cache_write FROM api_calls WHERE cache_read = 0 AND cache_write > 10000 ORDER BY ts DESC LIMIT 10"

"Why did the cache bust?"

Cache busts show as cache_read=0 with large cache_write. Common causes:

• Session switch: cron heartbeat ran between your turns (different session = different prefix)
• Model switch: changing models invalidates the entire cache
• Restart/deploy: new process = new cache
• Compaction: context was compressed, changing the prefix

"Can I afford this?"

Rough estimates (Opus, relative to mana):

• Simple Q&A turn: ~$0.07-0.10 (cache hit) → <0.1% mana
• Multi-tool investigation (5 turns): ~$0.50-1.00 → ~0.5% mana
• Long coding session (30 min): ~$5-15 → ~4-10% mana
• Cache bust + rebuild: ~$2.50 one-time → ~2% mana

How Mana Works

1. Anthropic tracks API usage in a 5-hour window that resets to 100%
2. Utilization = (usage in window) / (plan limit) x 100
3. Mana = 100% - utilization
4. Old usage rolls off the trailing edge — mana recovers over time
5. The window slides continuously, not in discrete steps

Why the ratio varies

At low utilization, the $/mana% ratio appears inflated. This suggests Anthropic's rate limiter isn't purely dollar-cost-based — it may factor in token counts, request rates, or have a non-linear curve. Above ~15% utilization, the ratio stabilises.

Mana data sources

• Mana readings: Embedded in [meta] headers in session JSONL files. Cached for 5 minutes by foci (fetched from Anthropic's OAuth usage API).
• API costs: Logged per-call in ~/data/api.db with cost_usd, token counts, session ID, timestamps.

Common Mana Drains

1. Cron heartbeats — each invocation runs multiple tool calls on Opus. Can cost several percent mana per hour even with no user activity.
2. Cache busts — session interleaving (cron fires between user turns), deploys, model switches.
3. Separate sessions — each session maintains its own cache prefix. A session that hasn't been active recently pays a full cache write on its first turn.