All recipes

$0.148 → $0.082per 3-task sequence: full CLAUDE.md in fresh sessions vs trimmed CLAUDE.md in one continued session

Does Claude Code cache your startup context? Yes, but fresh sessions pay for CLAUDE.md again

We ran the same 3 small tasks with a 1,501-word or a 9-line CLAUDE.md, in fresh sessions or one continued session, 3 runs each, and logged cache reads and writes on every API request. Every run's first request read the same 13,069 tokens from cache and sent 2 uncached. CLAUDE.md was another story: fresh sessions wrote it to cache again for each task. Trimming cut cost 20 to 22%; staying in one session cut 29 to 31%; both together, 45%.

• context
• cache
• claude-md
• cost

$2.64 vs $3.28average cost per run, Fable 5.2 (Claude Code 2.1.321 rerun) vs Fable 5.1 (both fixed the bug in every scored run)

Fable 5.2 is 20% cheaper per token. On our bug fix it averaged $2.64 a run, under Fable 5.1's $3.28

Launch-day test: Claude Fable 5.2, Fable 5.1 and Opus 5.5 on the same planted bug in undici (902 files), same prompt and tools. Fable 5.2 (rerun on Claude Code 2.1.321) fixed the bug in 3 of 3 and averaged $2.64 a run and 2.37M tokens. Fable 5.1 averaged $3.28 and 2.04M; Opus $0.77 and 1.39M. Small samples: 3, 2 and 3 runs.

• models
• cost

3 of 3 eachruns kept SAVE10 coverage and caught a 9% mutation; $0.013 a run for Haiku 5.5, $0.117 for Sonnet 5.5

Test cleanup is the job everyone wants to hand off. Can the cheap model do it without quietly deleting coverage?

We gave Sonnet 5.5 and Haiku 5.5 the same messy test suite: duplicates, skips, a dead import, and one precise SAVE10 assertion that looks redundant. Both kept the coverage and caught a SAVE10→9% mutation in 3 of 3 runs. Haiku averaged $0.013 a run; Sonnet $0.117, about 9× more.

• models
• testing
• cost

0 of 15runs that started a subagent, on a 902-file repo (all 15 fixed the bug)

We gave Opus a 902-file repo and a search-heavy bug. It started 0 subagents in 15 runs

A follower said our last task was too small for Opus to delegate. So we used undici, Node's HTTP client (902 files, pinned commit), and a bug that needs tracing through 4+ files. Sonnet alone, Opus with subagents off and Opus with default subagents all fixed it in 5 of 5 runs, and none of them started a single subagent. Sonnet cost $0.248 a run; Opus $0.776 with subagents available and $0.951 without.

• subagents
• models
• cost

$0.051 vs $0.073per run, a 9-line decisions-only CLAUDE.md vs an 84-line full description (all 9 runs avoided the trap)

Should your CLAUDE.md describe the code, or just the decisions? We tested 3 versions

Same task, same planted trap, 9 runs: no CLAUDE.md, a 9-line decisions-only file, and an 84-line full description. Sonnet avoided the trap and passed the hidden test in all 9. The decisions file was the cheapest at $0.051 a run; the full description cost $0.073, 43% more.

• claude-md
• project-setup
• cost

$0.30API cost for Claude Code to build it: 2 prompts, 126 seconds

We had Claude Code build a cost tracker. It took 2 prompts and $0.30.

Claude Code (Sonnet 5.5) built cc-cost, a small CLI that reads its own session logs and prints API spend per day and per session. It took 2 prompts, 126 seconds and $0.30. Its tests passed the first time, but the real logs turned up an undercount the tests missed. After one fix it matched Claude Code's own cost totals on 76 of 86 real sessions.

• builds
• cost

$0.649 vs $0.283per run, the full Opus + Sonnet + Haiku setup vs Sonnet alone (2 of 3 vs 3 of 3 passing)

Opus plans, Sonnet builds, Haiku searches: the mix cost 2.3x Sonnet alone

We tested the viral 'run Opus, Sonnet and Haiku at once' Claude Code setup against plain Sonnet on the same 3-step feature in a 164-file repo. Sonnet alone passed every check in 3 of 3 runs for $0.283 a run. The full mixed setup passed 2 of 3 for $0.649, and Haiku did 1.8% of the work.

• models
• subagents
• cost

0 forksin 9 runs with forking on by default; Claude did all the work itself

Claude Code forks subagents by default now. In 9 runs, it never forked

Claude Code turned on fork subagents by default in claude -p. On a 164-file repo with a three-step job, Claude never forked or delegated once in 9 runs. Forks on cost $0.218 a run, forks off $0.245, and with subagents denied $0.305.

• subagents
• context
• cost

14% cheaperto stay in one session on a 197-file repo; every hidden test passed both ways

/clear on a bigger repo: one session was still cheaper

We reran our /clear test on a 197-file codebase with 14 feature modules, up from 24 files last time. All 24 hidden-test checks passed both ways. Staying in one session cost 14% less ($0.137 vs $0.159 a run).

• context
• workflow
• cost

5 of 5 eachruns passed every hidden test; $0.054 and 18s a run for Sonnet 5.5, $0.127 and 148s for Kimi K3.1 Pro

Kimi K3.1 Pro vs Sonnet 5.5 in Claude Code: same score, Kimi cost 2.4x as much and took 8x as long

We ran the same tricky billing-date task 10 times in Claude Code, 5 with Claude Sonnet 5.5 and 5 with Kimi K3.1 Pro through Moonshot's Anthropic-compatible endpoint. All 10 passed every hidden test. Kimi checked its work more often, but averaged $0.127 and 148 seconds a run against $0.054 and 18 seconds for Sonnet.

• models
• cost

3 of 3runs passed every hidden test either way; only 'think hard' runs added and ran tests (3 of 3 vs 0 of 3)

Does 'think hard' actually help? Same pass rate, but it tested its own work

We asked Claude Code to implement a billing-date function full of edge cases six times. Three runs got the plain request, three also got "Think hard about this before you write any code." All six passed every hidden test. Only the think-hard runs added tests and ran them, for about 60% more per run.

• prompting
• testing

2 of 3runs fully fixed the bug with the stack trace pasted (0 of 3 with a description)

Paste the error, or describe it? Both found the bug; only pasted errors caught the second crash

We gave Claude Code the same crashing notes API six times. Three runs got a one-line description of the symptom, three also got the stack trace. All six found the root cause. Only the pasted-error runs also fixed a second crash in tag counts, in 2 of 3 runs, for about half a cent more per run.

• debugging
• prompting

4,271 → 522tokens the setup added to every session, before and after the audit

We had Claude Code audit a bloated setup. It cut 88% of what the setup loaded.

A 1,500-word CLAUDE.md and 8 skills added about 4,300 tokens to every session. After an audit, the setup added about 500 tokens and every rule still held. That was a cut of around 16% of startup context, not most of it.

• claude-md
• context
• cost
• project-setup

15 vs 18 of 18runs passed, local Qwen 27B (thinking on) vs Sonnet 5.5 (adaptive)

A 27B model on two home AMD cards passed 15 of 18. Sonnet passed 18.

We ran Qwen3.8-27B on two Radeon AI PRO R9700s against Claude Sonnet 5.5 and Haiku 5.5 on six auto-graded tasks. Sonnet with thinking passed all 18 runs for $0.30. Qwen with thinking passed 15 for $0 in API, and most of its git patches didn't apply.

• local-models
• models
• cost

73% morecost per run with plan mode first (6 of 6 passed either way)

Plan mode first, or just start? Both passed; planning cost 73% more

We asked Claude Code to add note archiving to a 24-file API six times. Three runs planned first in plan mode, three just started. All six passed our hidden tests. Planning first cost about 73% more per run, $0.28 against $0.16.

• plan-mode
• workflow
• cost

3 of 3runs passed the hidden edge-case test with tests first (2 of 3 without)

Ask for tests first: 3 runs with, 3 without

We asked Claude Code for a bill-splitting function six times. Asked plainly, it shipped a rounding bug in 1 of 3 runs. Told to write failing tests first, 3 of 3 passed our hidden edge-case test, for about 2 cents more per run.

• testing
• prompting

0 of 3runs edited the generated file with a CLAUDE.md (2 of 3 without)

What a short CLAUDE.md actually changes: 3 runs with, 3 without

We gave Claude Code the same task six times. Without a CLAUDE.md it edited a generated file in 2 of 3 runs. With seven lines of rules, 0 of 3, and it was cheaper too.

• claude-md
• project-setup

17% cheaperto stay in one session, even with unrelated tasks

Unrelated tasks, same answer on /clear

Our last test found staying in one session cheaper for related tasks. We repeated it with three unrelated tasks in different parts of a codebase. Staying still won on cost, with one catch.

• context
• workflow
• cost

28% cheaperto stay in one session than to clear between tasks

Is /clear worth it? Not for small, related tasks

Three small tasks in the same files, run in one long session or with a fresh session each time. Same results, but clearing cost more, because the long session's context was cached.

• context
• workflow
• cost

How to set up a Claude Project so it works from the first message

What to link, what to write in the instructions, and what to hand over on day one, written by a Claude that works inside one. Each step says what you do and what Claude needs from you.

• projects
• project-setup

4 of 4ways got the rule followed; one CLAUDE.md line was cheapest

Hook, skill or CLAUDE.md line? We ran all three (and nothing)

We asked Claude Code to follow a house formatting rule four ways, 3 runs each. All four worked. The cheapest was one CLAUDE.md line, and the hook cost more than we expected.

• skills
• hooks
• claude-md
• project-setup

24% morewith a subagent for searches, and no context saved

Subagents for codebase searches cost more and saved nothing

A three-step refactor in a 24-file API, with and without a CLAUDE.md line telling Claude to delegate searches to a subagent. Both got everything right. Delegating cost more, at least at this size.

• subagents
• context
• cost
• claude-md

$0.009 vs $0.221per run, Haiku vs Opus, same 9 of 9 passes

Haiku was enough

We gave all three models a checkout feature with coupons, tax and edge cases. All nine runs passed every hidden test. Haiku averaged under a cent; Opus cost 25 times more.

• models
• cost