Free tool · Planner
AI Budget Planner
What will your AI stack cost over the next six months? Put API usage, connected MCP tools, agents and subscription seats in one plan, set your growth, and see every month before it arrives. The tool line is the one most teams miss: every connected tool is re-sent on every call.
Next 6 months. About $13,163 over 6 months on Claude Sonnet 5.5.. Month 1 is $1,719; month 6 is $2,732. The largest share is connected tools.
Next 6 months
$13,163
$2,194 a month on average; month 6 runs at $2,732.
- API usage$6555% of the total
- Connected tools$11,09784% of the total
- Agents$1,0518% of the total
- Subscriptions$3603% of the total
API usage
Cached share bills at 10% of the input price, Anthropic's cache-read rate on most Claude models. Thinking tokens bill as output, so count them in the reply size.
Connected tools (MCP)
Every request that carries tools sends their names, descriptions and schemas as input tokens. With this setup that is 47.3K tokens on each call, including the 286-token tool prompt Anthropic adds for Claude Sonnet 5.5. Agent steps carry them every step.
Agents
With context building up, every step re-sends the prompt plus everything said and returned so far. That is why step count moves the bill more than run count.
Subscriptions
Monthly list prices from our subscription table, only plans whose price we read on a known day. Annual billing is often cheaper.
What would move your total most
Each line doubles one input and keeps the rest. Firm up the top guess first.
- A model at twice the price+$12,803
- Connected tools doubled+$11,030
- Agent steps per run doubled+$9,355
- API requests per day doubled+$4,650
- Reply length doubled+$933
An estimate at today's list prices, not a quote. Not included: future price changes, the one-off cache-write premium, long-context surcharges, and per-call fees of server tools such as web search.
The data points behind it
Tool definitions are billed input
Tool names, descriptions and schemas count as input tokens on every request that includes them.
Anthropic: Tool use with Claude, Pricing, read 2026-10-09Claude's tool-use system prompt
Anthropic lists 286 to 497 tokens per request when tools are present, for the Claude models this planner prices.
Anthropic: Tool use with Claude, Pricing, read 2026-10-09Measured MCP server sizes
58 tools across five servers took about 55K tokens before the conversation started.
Anthropic Engineering: Introducing advanced tool use (24 November 2025), read 2026-10-09Loading tools on demand
Anthropic reports an 85% reduction in tool tokens with tool search.
Anthropic Engineering: Introducing advanced tool use (24 November 2025), read 2026-10-09Cached prompt reads
A cache read bills at 10% of the base input price on most Claude models.
Anthropic: Prompt caching, read 2026-10-09Model and seat prices
Per-token prices come from our model price ledger, read 2026-09-03. Seat prices come from our subscription table, oldest price read 2026-09-12.
Frequently asked
How much does an AI stack cost over six months?
It is the sum of four lines that grow differently: API usage and agent runs grow with your traffic, connected tools add a fixed number of tokens to every call that carries them, and subscription seats stay flat. Pick an example above, change the inputs to your own, and the planner projects each month with your growth rate. The total for 6 months is the figure to budget against; the month 6 number is what you will be paying by the end.
How many tokens do MCP tools add to each request?
Every connected tool's name, description and input schema is sent as input tokens on each request that carries it. Anthropic measured GitHub about 26K, Slack about 21K, Jira about 17K, Sentry about 3K, Grafana about 3K, Splunk about 2K, and a five-server setup of 58 tools at about 55K tokens before the conversation starts. That works out near 948 tokens a tool, the planner's default for tools not in that list. On Claude models the API also adds a tool-use system prompt of 286 to 497 tokens, depending on the model.
How do I cut the cost of connected tools?
Three ways. Load tools on demand: Anthropic reports an 85% reduction in tool tokens with its tool search feature. Cache the tool list, since it rarely changes between calls; cached reads bill at 10% of the input price on most Claude models. And connect only the servers a task needs, because agents re-send the full list on every step.
Why do agents cost so much more than chat?
An agent is a loop: each step sends the prompt plus everything said and returned so far, so input grows with every step. The tool list rides along on each of those steps too. Doubling steps per run usually moves the bill more than doubling runs, which the planner's 'What would move your total most' list shows for your own inputs.
How accurate is the estimate?
It is as accurate as your inputs, at today's list prices. Model prices come from our price ledger (read 2026-09-03) and seat prices from our subscription table (oldest price read 2026-09-12). It leaves out future price changes, the one-off cache-write premium, long-context surcharges and per-call fees of server tools such as web search. Check your provider's usage dashboard after the first month and replace the guesses with real numbers.