For teams running long agent sessions
Does prompt caching actually cut what our agents cost?
Short answer: Often a lot, for agents that re-read the same instructions, files and history every turn. Anthropic prices cache reads at a fraction of normal input, but common actions like switching models or adding a tool server break the cache and make you pay full price again. Check your cache hit numbers before assuming it works. Start with the $150 burn-audit hour.
How it saves money
Agents resend their instructions, files and conversation on every turn. Caching lets the provider reuse what it already processed. Anthropic’s prompt caching docs price a cache read at 0.1× the base input price on most models (lower on some), while writing to the cache costs more than a normal read: 1.25× for the 5-minute cache and 2× for the 1-hour cache. So caching pays off when the same prefix is reused quickly, and wastes money when it isn’t.
What breaks the cache
Claude Code’s prompt caching docs list actions that invalidate the cache, including switching models, changing effort level, connecting or removing an MCP server, enabling or disabling a plugin, and compacting the conversation. Each one means the next turn re-reads everything at the normal rate. An agent that switches models mid-task, or a setup that adds tools on the fly, can quietly lose most of the saving.
Start with the $150 burn-audit hour. One hour with an operator. You leave with a written map of the jobs, tokens and flows to cut or rebuild. Pay the hour, then we schedule the session. See a sample audit (example data) or try the free burn check.
Book the hour — $150 Operator pack — $625 depositAI Concierge — $2,000/mo
The pack is $1,250 for 10 hours; checkout takes the $625 deposit. Questions first? Email daniel@agenthiveinc.com or call our AI front desk demo at (320) 335-6186.
How to check yours
- Look at the cache read and cache write token counts your tool or API reports per turn. A high read-to-write ratio means caching works.
- Find sessions where writes stay high turn after turn: something in the prefix keeps changing.
- Keep stable instructions at the start and changing content at the end.
- Pick the model and tools before a long task, not halfway through.
Prices and multipliers change by model, so confirm on Anthropic’s API pricing page.
The three ways to work with us
- $150 burn-audit hour. One working hour. We map the agents, routines and flows your team runs and hand back a written list of what to cut, fix or rebuild.
- $1,250 operator pack ($625 deposit). Ten hours of build and fix work on the flows from the audit. The balance is due as the work is scheduled.
- AI Concierge, $2,000/mo. A senior AI operator on retainer: two 45-minute working sessions a month, unlimited async help between them, and a shared asset inventory. Month to month; cancel any month after the first 30 days.
Common questions
Is caching on by default?
It depends on the tool. Claude Code uses it automatically; with the API you may need to turn it on. Check your provider’s docs.
Can caching make costs go up?
Yes, if content is written to the cache and never read back before it expires, because writes cost more than normal input.
Does this replace trimming context?
No. Caching makes re-reading cheaper; trimming means less to re-read. Do both.
Start with the $150 burn-audit hour. One hour with an operator. You leave with a written map of the jobs, tokens and flows to cut or rebuild. Pay the hour, then we schedule the session. See a sample audit (example data) or try the free burn check.
Book the hour — $150 Operator pack — $625 depositAI Concierge — $2,000/mo
The pack is $1,250 for 10 hours; checkout takes the $625 deposit. Questions first? Email daniel@agenthiveinc.com or call our AI front desk demo at (320) 335-6186.
Related
- Why do agents that re-read long histories burn so much usage?
- Why do sub-agents and multi-agent setups burn so many tokens?
- How do we cut AI agent spend without breaking the work that matters?
- Free AI agent burn check
Outside facts link to their sources and were checked in November 2026; product pricing and limits change, so confirm on the vendor’s page. First Deploy is built and run by Daniel Graham at AgentHive Inc, Palm Coast, Florida. We don’t promise a savings figure before we’ve seen your setup.