Task Budgets & Token Economics — Models & Architecture
[ ABORT TO HUD ]
SEQ. 1
SEQ. 2
SEQ. 3
Task Budgets & Token Economics
Managing Token Budgets in Series 5.5
In high-throughput and long-horizon agentic environments, mastering Anthropic's October 2026 rate card and cost modifiers is critical:
- Claude Fable 5.1: $10.00 / MTok input ($0.25 / MTok cache read at 0.025x), $50.00 / MTok output.
- Claude Opus 5.5: $4.00 / MTok input ($0.20 / MTok cache read at 0.05x), $20.00 / MTok output.
- Claude Sonnet 5.5: $2.00 / MTok input ($0.10 / MTok cache read at 0.05x), $10.00 / MTok output.
- Claude Haiku 5.5: $0.10 / $0.50 per MTok (≤100K prompt; $0.01 cache read at 0.10x) and $0.50 / $2.50 per MTok (>100K prompt; $0.05 cache read at 0.10x).
- Prompt Cache Write: 1.25x base input rate for default 5-min TTL, or 2.0x base input rate for 1-hour TTL (
"ttl": "1h"); 512-token minimum prefix across 5.x. - Batch API: Flat 50% discount on both input and output tokens across all models for asynchronous 24-hr SLA workloads.
- Data Residency (
inference_geo: "us"): 1.1x multiplier (10% premium across input, output, cache writes, and cache reads) when pinning inference strictly to US infrastructure vs default"global".
⌨ HANDS-ON LABAudit Token Budgets & Input Limits
⭐ +150 XPProtect against unbounded costs. Calculate exact token consumption on Claude Sonnet 5.5 using count_tokens before making downstream calls.
1POST an oversized context payload to /v1/messages/count_tokens to measure exact input tokens before spending.
OBJECTIVE 1 / 1 — type "hint" if stuck
SYNAPSE VERIFICATION
QUERY 1 // 2
What is the cost reduction achieved on Prompt Cache Reads across Claude 5.x models?
10% discount
50% discount
90% to 97.5% discount (0.1x on Haiku 5.5, 0.05x on Sonnet 5.5 & Opus 5.5, and 0.025x on Fable 5.1)
100% free