Five levers that move your coding-agent bill

Lever 01

Model and reasoning

Which model runs the task, and how much reasoning effort it spends. Reasoning tokens are billed as output, the most expensive kind.

The number: xhigh reasoning effort went from 70% of spend in July to 22% in September. In a side-by-side run of the same prompt, Opus cost 5.1 times as much as Haiku, and $0.21 of the $0.33 gap was cache writes.

What to do Monday: Send routine work, such as renames, small fixes, and test runs, to a cheaper model at medium effort. Keep the top model and high effort for tasks where the path is unclear or a mistake is expensive. If you switch models partway through a task, start a fresh session with a short handoff note. The new model can't reuse the old model's cache, so carrying the whole history over means writing all of it again.

Where to see it: Models groups each model by reasoning effort. Sessions → Compare puts up to four runs side by side, with model, effort, and cost split by token type.

Lever 02

Task context

Everything loaded before the agent starts working: instructions, tool and MCP definitions, skills, files, and earlier turns. You pay for it on every call.

The number: 83,907 tokens of tool definitions were loaded into prompts and never called. Separately, subagent cost per run fell from $6.04 in August to $0.96 in September, as runs became shorter and more narrowly scoped.

What to do Monday: Turn off MCP servers and skill packs a project doesn't use. Give each subagent one specific job and only the files it needs. When most of a session's context is history the current step no longer needs, compact it or start fresh.

Where to see it: Tools lists definitions that were loaded and not called. Treat “no use observed” as a prompt to review, not proof that a tool is useless. Efficiency tracks context load, and Subagents shows cost per run for each role.

Lever 03

Prompt caching

Providers store a repeated prompt prefix so later calls can read it cheaply. Writing to the cache costs more than plain input, so a cache only pays off when the prefix stays the same long enough to be read many times.

The number: cache writes were 55% of the Opus 5.5 bill.

What to do Monday: Keep the start of the prompt stable. Avoid editing instruction files or changing the tool list in the middle of a session. Put material that changes often near the end. Cache entries also expire after a few idle minutes (five by default on Anthropic's API), and the next call writes them again, so a session left idle and resumed many times pays for repeated writes.

Where to see it: each session shows cache writes beside cache reads. Efficiency tracks the cache hit ratio over time.

Lever 04

Outcome-focused prompts

Telling the agent what finished looks like, instead of describing activity.

The number: 15% of 1,964 follow-up messages were pushback: corrections, redirections, and requests to redo work.

What to do Monday: Name the outcome, limit the scope, and say how to test it. For example:

Fix the failing date test in tests/test_dates.py.
Change only the date parser module.
Done means the test suite passes with no other test changes.

Where to see it: run the same task once with your usual prompt and once with an outcome-focused one, then open both in Sessions → Compare. Executions, active time, and cost sit next to each other.

Lever 05

Avoid rework

Rework is the expensive part of a session. Each correction in a long session re-reads the whole context before the agent writes a word.

The number: sessions that recovered after pushback cost $33 on average. Sessions accepted without pushback cost $6.

What to do Monday: If a session goes off track twice, stop pushing. Start a new one with a short handoff: the goal, what was tried, what failed, and the next step. For larger changes, review the agent's plan before it writes code.

Where to see it: sort Sessions → All sessions by cost and open the most expensive runs to see executions and peak context. Token Meter doesn't decide whether a session succeeded. That call is yours.