Harness experiment · 3 agents · 1 CRM · October 2026
Claude Code vs Codex vs Pi: Two Days Building One CRM
Three AI coding harnesses building one CRM on one laptop, measured two days in. Planning finished on Oct 28; then Claude Code and Codex started building, and Pi joined on Oct 29. How long had each been running, how much of that was real work, and what would it cost?
Data as of Oct 30, 06:35. Pipeliner, a local CRM: companies, people and opportunities on four pipelines.
Claude Code
Snapshot, two days in
41.0h running, 28.5h active
Started Oct 28, 13:38, right after the plan review. Nonstop since Oct 29, 03:56: 25.9h active of the last 26.6h.
- Agent-hours, helpers added
- 102.2h
- Blocked by usage limit
- 5.5h
- Waiting for you
- 6.5h
- v1 tagged after
- 39.2h (26.8h active)
- Cost at API prices
- $1,579 est.
- Test functions
- 1,718
Model: Claude Opus 5.5 (Sonnet 5.5 for some helpers)
Codex
Snapshot, two days in
38.7h running, 20.7h active
Started Oct 28, 15:53. Built v1 in under two hours, then the MVP fix plan and fix cycles.
- Agent-hours, helpers added
- 50.5h
- Blocked by usage limit
- 0h
- Waiting for you
- 16.1h
- v1 validated after
- 1.8h (1.8h active)
- Cost at API prices
- $519 est.
- Test functions
- 677
A second Codex build ran in parallel on Oct 28 and was stopped: 2.8h active, about $68.32, nothing kept. Model: gpt-6-astra and gpt-6.1-sol
Pi
Snapshot, two days in
19.2h running, 7.4h active
Added Oct 29, 11:20, on top of Codex's code. No build from scratch: it iterates.
- Agent-hours
- 7.4h
- Blocked by usage limit
- 0h
- Waiting for you
- 10.8h
- MVP steps 3 to 8 done after
- 6.2h (2.2h active)
- DeepSeek bill (dashboard)
- $24.34 real
- Test functions
- 483 (+84)
Model: DeepSeek V4 Pro
Claude built the whole plan from an empty repo, with checks after every lane and wave. Codex built the same plan in under two hours on top of the foundation Claude wrote first. Pi never built; it iterates on Codex's code, so it appears only in the iteration comparison.
Running versus active
From each build's start to the snapshot: how long it had been running, and how much of that it was actually working.
| Claude Code | Codex | Pi | |
|---|---|---|---|
| Started | Oct 28, 13:38 | Oct 28, 15:53 | Oct 29, 11:20 |
| Running for (to the snapshot) | 41.0h | 38.7h | 19.2h |
| Active | 28.5h | 20.7h | 7.4h |
| Active share | 70% | 53% | 38% |
| Agent-hours (helpers added) | 102.2h | 50.5h | 7.4h |
| Idle: blocked by usage limit | 5.5h | 0.0h | 0.0h |
| Idle: waiting for you | 6.5h | 16.1h | 10.8h |
| Idle: asleep or other | 0.4h | 2.0h | 1.0h |
| Cost | $1,579 at API prices | $519 at API prices | $24.34 billed |
Where the running time went
Each build's running time, split into active work and the reasons it was idle.
WorkingBlocked by usage limitWaiting for youLaptop asleepOther idle
When each one worked
Minutes of work in each hour, main session and helpers together.
The v1 build
Claude and Codex each built all of plan.md. Time to their v1 done point, with the parallel Codex build that was stopped.
| Claude Code | Codex | Codex, stopped | |
|---|---|---|---|
| Plan | plan.md v1, all phases | plan.md v1, all phases | plan.md v1, all phases |
| Started from | Empty repo | Claude's foundation (172 tests, 122 KB) | Claude's foundation |
| Start | Oct 28, 13:38 | Oct 28, 15:53 | Oct 28, 15:45 |
| Done | Oct 30, 04:52 (v1.0.0 tag) | Oct 28, 17:38 (validated) | Stopped Oct 29, 19:40 |
| Clock time | 39.2h | 1.8h | 27.9h |
| Working | 26.8h | 1.8h | 2.8h |
| Agent-hours (helpers added) | 93.7h (285 helpers) | 6.4h (26 helpers) | 11.2h (4 helpers) |
| Idle: usage limit | 5.5h | 0.0h | 0.0h |
| Idle: waiting for you | 6.5h | 0.0h | 3.5h |
| Idle: asleep or other | 0.4h | 0.0h | 21.6h |
| Cost at API list prices | $1,468 | $54.22 | $68.32 |
| Test functions at done | 1,718 | 334 (+162) | None committed |
| Python in crm/ at done | 1,857 KB | 494 KB (+372) | None committed |
| Checking | Check, code review and verify per lane; review per wave | One validation pass at the end |
Iteration, for reference
Not part of the build: MVP fix plan steps 3 to 8, done by Codex and by Pi from the same code.
| Codex | Pi | |
|---|---|---|
| Started from | Codex's Oct 29 09:00 commit (399 tests) | The same commit |
| Clock time | 11.3h | 6.2h |
| Working | 8.9h | 2.2h |
| Agent-hours | 19.6h | 2.2h |
| Cost | $148 at API prices | $6.29 recorded by Pi |
| Test functions added | +182 | +65 |
| Python added in crm/ | +173 KB | +56 KB |
| Checking | verify after every step | Tests and lint only, no verify |
What we learned
- Claude never stops on its own. 41.0h running, 28.5h active, and nonstop since Oct 29, 03:56 (25.9h of the last 26.6h). Its only idle time was its 5-hour usage limit (5.5h) and the first night, waiting for you to say "continue" (6.5h).
- Codex and Pi wait for you. Codex was active 53% of its 38.7h and Pi 38% of its 19.2h. Most of their idle time (16.1h and 10.8h) was waiting for your next message.
- Parallel agents multiply the hours. Claude's 28.5h active held 102 agent-hours of work across 285 helpers. On the first afternoon about 20 agents ran at once and used up the 5-hour window twice in four hours.
- Fast versus thorough. Codex built v1 in 1.8h for about $54.22, on top of the foundation Claude wrote in its first 90 minutes, with one validation pass. Claude took 26.8h of work and about $1,468, checked and reviewed every lane and wave, and finished with 1,718 test functions to Codex's 334.
- Two Codex builds ran at once by accident. The one you stopped used 11.2 agent-hours and about $68.32, and left nothing behind.
- Iteration is a different game. On the same plan steps from the same code, Pi worked 2.2h for $6.29; Codex worked 8.9h for about $148, checked every step and added nearly three times the tests.
- Subscriptions hide the price. At API list prices Claude's run to the snapshot would cost about $1,579 and Codex's about $519; 51% of Claude's is re-reading cached context. Pi's DeepSeek bill is $24.34. Neither the Claude Console nor OpenAI's API dashboard shows this work: both harnesses ran on subscriptions.
- No quality verdict yet. This measures build time, effort, output and cost. Which build works best on real mail and real leads is the next test.
Four changes to the process, one for each stage, in one repo: github.com/DoryZi/token-efficient-lifecycle.
How we measured
- Running time: from the message that asked each harness to build (Claude Code Oct 28 13:38, Codex Oct 28 15:53, Pi Oct 29 11:20) to the snapshot on Oct 30, 06:35. Claude's session opened at 09:02 on Oct 28 for the plan review; counted from there it is 45.6h running and 29.3h active. Counted: each harness's build session and its helpers. Not counted: planning, and separate side sessions (the feedback bubble tool, local names, these reports). The v1 section ends at each build's done point.
- Working time: a log moving with gaps of 10 minutes or less, plus every tool call for its full length (capped at 60 minutes per call), plus background test runs and agents from launch to completion. Servers, watchers and the plan deck do not count. Main session and every helper, merged; agent-hours add them all up.
- Idle time: blocked by usage limit (from Claude's 'You've hit your 5-hour limit' to the reset it names), waiting for you (idle that ended with your message, or time waiting on a plan approval or question), laptop asleep (system journal), other.
- Claude's helpers include its workflow agents. Codex's include the agents it spawned. The stopped Codex build is the session started at 15:45 on Oct 28 with its three agents, stopped at your request on Oct 29.
- Cost: every model call's tokens times the API list price (Claude Opus 5.5 $4 in, $20 out, $0.20 cache read, $5 or $8 cache write; gpt-6.1-sol $2 in, $0.10 cached, $10 out). Pi records its own DeepSeek cost; DeepSeek's dashboard billed $24.34 for 653M tokens over 30 days, against $31.92 recorded by Pi.
- Test functions are def test_ lines in the code at the done commit, not tests that passed.