SAT, SEPTEMBER 26, 2026
Independent · In‑Depth · Practitioner‑Tested
Claude Productivity

What Your Agents Actually Cost: Seven Cost-Modelling Prompts

Anthropic just committed $11.6 billion to Akamai for CPU capacity, not GPUs. That is the frontier telling you the bill around the model is real and large. Meanwhile DeepSeek raised prices 2.3x to 4.5x in August and OpenAI halved its own in September, so any cost model built before this month is wrong. These seven prompts rebuild it from your actual logs rather than from a pricing page.

⌨️ 7 prompts 🕐 Updated Sep 24, 2026
💡 How to use these prompts: Replace everything in [BRACKETS] with your specific details before sending. Click Copy to copy any prompt to your clipboard instantly.
1
Cost Per Completed Task, Not Per Million Tokens
A model at half the price that needs three attempts is more expensive. This is the only comparison that settles it.
Below is a sample of my production API logs.<br/><br/>Do not tell me cost per million tokens. Tell me cost per COMPLETED task.<br/><br/>Calculate:<br/>1. Total spend across the sample<br/>2. Number of tasks that reached a successful end state<br/>3. Spend divided by successful tasks<br/>4. What share of spend went to runs that failed, were retried, or were abandoned<br/><br/>Point 4 is the number I actually want. Show your working.<br/><br/>LOGS:<br/>[paste]
2
Reprice Against Today's Rates
DeepSeek rose 2.3-4.5x in August, OpenAI fell 50% in September. Any July decision needs redoing.
Here is my token usage for the last 30 days, broken down by model.<br/><br/>Reprice it against current list rates for each of: GPT-6 Luna, GPT-6 Sol, GPT-6 Astra, Claude Opus 5.5, DeepSeek V4.1 Flash at peak and off-peak, and Grok 4.7.<br/><br/>Output a table sorted by total cost, cheapest first, with a column showing the difference against what I actually paid.<br/><br/>Note explicitly where cached input rates change the ranking, because on some workloads they do.<br/><br/>USAGE:<br/>[paste]
3
Find the Cache You Are Not Using
Cached input is 10-100x cheaper than fresh input. Most pipelines pay full rate for text they send every time.
Analyse the prompts in these logs.<br/><br/>Identify every block of text that is repeated across many requests and is a candidate for prompt caching - system prompts, tool definitions, reference documents, few-shot examples.<br/><br/>For each: how many characters, how many times it was sent, and what it cost at full input rate versus what it would cost at cached rate.<br/><br/>Rank by money left on the table.<br/><br/>LOGS:<br/>[paste]
4
The Bill Around the Model
Anthropic just committed $11.6bn to CPU capacity. The non-token bill is not a rounding error at scale.
I want the total cost of running this agent, not just the API line.<br/><br/>From the architecture below, list every cost that is not tokens: compute for the orchestration layer, sandbox or container time, vector database queries, storage, egress, monitoring, and anything else you can identify.<br/><br/>For each, say whether it scales with number of runs, length of run, or neither.<br/><br/>If I cannot estimate one from what I have given you, say so rather than guessing.<br/><br/>ARCHITECTURE:<br/>[paste]
5
Route by Task, Not by Habit
Most pipelines run one model for everything. Routing is usually the largest single saving available.
Below is a list of the distinct task types my agent handles, with rough volumes.<br/><br/>For each, recommend the cheapest model that will actually do it, and state the specific benchmark or capability reason.<br/><br/>Rules:<br/>1. If a task needs tool calls in a loop, do not recommend a budget tier - say why<br/>2. If a task is fully specified with no decisions, recommend the cheapest thing that handles it<br/>3. Give me the projected monthly cost of the routed setup versus running everything on one model<br/><br/>TASKS:<br/>[paste]
6
Find the Retry Waste
Retry waste is invisible on a pricing page and frequently double digits as a share of spend.
Search these logs for every request that was retried, and every run that consumed tokens without reaching a result.<br/><br/>Group by cause: rate limit, timeout, malformed output, tool error, model refusal, or unknown.<br/><br/>For each group: count, total tokens burned, cost.<br/><br/>Then tell me which single cause, if fixed, removes the most spend.<br/><br/>LOGS:<br/>[paste]
7
Stress the Model at 10x
A provider that has repriced 4.5x once can do it again. Model it before it happens, not after.
Here is my current monthly agent spend and volume.<br/><br/>Project the bill at 10x current volume under three scenarios:<br/>1. Prices stay flat<br/>2. My primary provider raises prices 3x, as DeepSeek did in August<br/>3. A cheaper competitor appears at half my current rate and I migrate<br/><br/>For scenario 2, tell me at what point the migration cost is worth paying.<br/><br/>State your assumptions explicitly so I can argue with them.<br/><br/>CURRENT:<br/>[paste]