Claude Productivity

Five Prompts to Price a Task Instead of a Token

Two price moves in two days, and neither is readable from a per-token table. Claude Sonnet 5.5 kept identical pricing while cutting cost per task up to 30% through fewer tokens and batched tool calls. GPT-6.1 Sol claims parity with Astra on coding and computer use at one fifth the token cost, launched beside a $500-a-month tier whose speed mode charges up to 6x. The only number that settles any of it is what one finished task costs you. These five prompts work it out.

⌨️ 5 prompts 🕐 Updated Sep 30, 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.

Define What Counts as a Completed Task

Most cost comparisons fail here, before any measuring happens. Resent context is the usual omission.

I want to measure what one unit of work actually costs me across different AI models. Here is what my system does: [describe the workflow]. Help me define one "completed task" precisely enough to measure: where it starts, what has to be true for it to count as done, and what counts as a failure rather than a slow success. Then list everything that consumes tokens between those two points - including retries, tool call round trips, system prompts re-sent on every turn, and context that gets resent as a conversation grows. Tell me which of these I am most likely to forget to count.

Account for the Retries

The break-even success rate is the number that decides cheap-model-with-retries against expensive-model-first-time.

Below are logs from my AI workflow, including failed and retried attempts. Calculate: average tokens per successful task including all retries that preceded it, the share of total spend going to attempts that produced nothing, and how much a 10% improvement in first-attempt success would save. Then tell me whether a more expensive model with a higher first-attempt rate would be cheaper overall, and what success rate it would need to break even. LOGS: [paste]

Count the Tool Call Overhead

This is where Sonnet 5.5's "up to 30% less" comes from. Whether you see it depends entirely on how agentic your workload is.

My agent workflow makes tool calls. Here is a trace: [paste]. Work out how much of my token spend is the actual work versus the overhead of calling tools - the schemas, the results being read back into context, and the context resent on each subsequent turn. Then tell me how much would change if the model batched calls instead of making them one at a time, and which of my tool calls could be batched without changing the result.

Price the Wait

For a batch job overnight, a 6x speed premium buys nothing. For a loop a developer is watching, it can pay for itself before lunch.

I am considering a premium speed tier that generates roughly [X] times faster at [Y] times the token cost. Help me work out whether it pays. I need to account for: how long someone is actually blocked waiting, what that time is worth, whether the wait is on a critical path or in the background, and how many times a day this happens. Give me the break-even - how many blocked minutes a day justify the multiplier - and tell me honestly which parts of my workload would see no benefit at all. WORKLOAD: [describe]

Build the Comparison That Would Actually Settle It

Deciding the switching threshold before the data arrives is what stops you rationalising whichever result you get.

I want to compare [model A] and [model B] on my own workload rather than on published benchmarks. Design the test: how many runs for a meaningful result, which tasks to use and why those, what to hold constant, and what to measure besides cost - quality, first-attempt success, latency, failure modes. Give me the table to fill in, with cost per completed task as the headline column. Then tell me what result would make me switch, written now, before I see the numbers. WORKLOAD: [describe]