AI Agent Cost Calculator
Estimate your agent’s cost per task and per month, compare models, and see what drives the bill.
Each one fills in the fields below for a typical workload of that kind. You can change any of them.
18 model calls · 1 sub-agent · caching on, 90% of input read from cache
Presets are illustrative starting points. Replace assumptions with your observed usage for a more accurate estimate.
Changing or resetting a preset replaces workload assumptions only. Usage and budget stay unchanged.
One task is one complete agent run, including its calls and any sub-agents.
Current model: Claude Sonnet 5.5
Annual spend: $41,772.85
Model API costs only. Excludes hosting, external tools, storage, subscriptions, and taxes.
Monthly budget (optional)
Advanced assumptions
If task lengths vary a lot, estimate light, normal, and heavy tasks separately. One average undercounts a few very long runs.
Other input is conservatively estimated at the five-minute cache-write rate, $2.50 per 1M. Cache eligibility, lifetime, and storage are not simulated. Short or cold-start runs may achieve less caching than long-running agents.
Estimate details
Input tokens per task: 897K. Output tokens per task: 17K.
One-model-call baseline: $0.011 using the entered system, new-input, output, and cache settings. The task is about 44× its request cost. This is a request-cost comparison, not an equivalent-work comparison.
How this is calculated
Retained history means later model calls resend earlier new input and output. For a whole loop, input is n × (system + new input) + (new input + output) × n × (n − 1) / 2.
Retries model the expected share of nominal calls that adds one retained-history recovery attempt. A fractional expected attempt is linearly weighted between adjacent whole loops.
With caching off, all input is ordinary input. With caching on, the share read from cache uses the read rate and the rest uses the model’s cache-write rate where it lists one, a conservative approximation. On at 0% means writes that are never reused.
Sub-agents start fresh independent loops using the selected assumptions.
Cost impact of changing one assumption
On Claude Sonnet 5.5, the current model.
| No retained transcriptTheoretical comparison, not a measured compaction saving. Quality and summarization costs are not modeled. | $1,274.37/month lower (36.6%) |
| No sub-agentsSub-agents per task: 0. The work they did is not replaced in this comparison. | $877.63/month lower (25.2%) |
| No retriesRetry rate: 0%. Assumes every call succeeds the first time. | $352.26/month lower (10.1%) |
| Current monthly spend | $3,481.07 |
Each scenario is calculated separately. Savings overlap and cannot be added. These are not equal-quality changes: removing history, retries, or sub-agents can change what the agent achieves.
Model prices
Claude Sonnet 5.5, $2.00 per 1M input, $10.00 per 1M output, $0.100 per 1M cached input, $2.50 per 1M written to cache. Verified 2026-10-30. Only models with a verified price are selectable. Displayed USD provider list prices only; custom rates are not accepted.
Pricing source (last verified 2026-10-30)Prices verified 30 Oct 2026. Public list prices in USD, before any discount you have negotiated.
Want to see what your agents actually spent?
Finitizer Tokenomics: Reads your AI bills and exports and shows what changed, who spent it, and what to fix. No proxy, no code change. Founding customers list now open.
Questions
Why can an AI agent cost more than a single API call?
A task can make several model calls. When history is retained, later calls can resend earlier input and output. Sub-agents and recovery attempts also affect the estimate.
What is one task?
One task is one complete agent run, including its calls and any sub-agents. For a customer-serving workload, choose Daily task volume and enter its total daily tasks.
How are caching and retries estimated?
With prompt caching off, all input is priced as ordinary input. With it on, the share of input tokens read from cache uses the read rate and other input uses the model’s listed cache-write rate where it has one, a conservative approximation. Retries represent the expected share of nominal calls that add one recovery attempt, not retry-until-success.
What does this estimate exclude?
It estimates displayed USD provider list-price model-token charges. It excludes subscriptions, external tool fees, compute, storage, taxes, and platform, regional, negotiated, or inference-mode differences. Custom rates are not accepted.
Related tools
LLM Token Cost Calculator
Compare monthly AI spend on Claude, GPT, Gemini and more before the invoice arrives.
OpenCloud Run Cost Calculator
Two billing models, and how busy your instances are picks the winner.
OpenAI Carbon Calculator
The carbon of AI prompts and GPUs, from the figures the makers have published.
OpenMore tools for: AI and model usage
A model is a forecast. Your invoice is the measurement.
Finitizer tracks real token consumption per agent, feature, and team so you can compare estimates with observed usage.
See Finitizer Tokenomics