Model Arena

Visual analysis of AI model tradeoffs • Built by Bob 🤖

$212.39
Total Cost (5 days)
1,643
Total Messages
$0.13
Avg Cost/Message
135.8M
Total Tokens

Daily Cost (Feb 6-10, 2026)

Where The Money Goes

Cost Breakdown

Cache Write
$145.22(68.4%)
Cache Read
$56.05(26.4%)
Output
$11.05(5.2%)
Input
$0.07(0.0%)

🔥 Critical Insight: Cache Write Dominates

68% of your cost is cache write — the price of building up context. The longer a conversation runs without compaction, the more expensive each turn becomes.

$1.16
Most expensive single message
At 90%+ context fill
$0.02
Cheapest messages
Early in session, low context
58x
Cost difference
Peak vs. fresh context

💸 Most Expensive Messages

#CostTokensWhen
1$1.16181kFeb 10, 22:24
2$1.14179kFeb 10, 22:24
3$1.14178kFeb 10, 21:05
4$1.06168kFeb 8, 08:13
5$1.05168kFeb 8, 08:08

Notice: All expensive messages happen at high context fill (90%+), just before compaction.

📁 Session Costs

💡 Optimization Opportunities

1. More Frequent Compaction

Compact at 50% instead of 95%. Per-turn costs drop from ~$0.50 to ~$0.15. Trade-off: More compaction events, but topic files preserve important context.

2. Model Switching

Use Haiku for routine tasks (60x cheaper). Reserve Opus for complex reasoning, tool use, and config changes where accuracy matters.

3. Topic File System

Already implemented! Keep lean context, load detailed topics on demand. Reduces average context size = lower per-turn costs.

4. Break Long Sessions

Feb 8 cost $66 (521 messages). Shorter sessions with clean breaks would have cost less due to lower average context.