Compare Claude Code and Codex by task outcome, evidence and cost

The data below is a synthetic demo workspace — invented projects and pricing-table cost estimates, never billed figures. Every number, evidence tier, provenance label and gap is the real receipt engine's own output; only the underlying work is invented.

The same task — "Add retry-with-backoff to the payments HTTP client", from the same starting commit — was run once on Claude Code and once on Codex. agentacct turns each into one Work Receipt. Read them side by side: not by whose summary sounds more confident, but by what actually happened, how well it is proven, and what it cost.

Claude Code

Work Receipt — Add retry-with-backoff to the payments HTTP client

task_cc_retry

• Decision status — VERIFIED · asserted by machine Recorded machine evidence verifies the latest outcome.
• Evidence coverage — 1/2 self-checked · 1 unchecked 1 non-verifiable (research/docs) Counts are passing checks over checkable steps, split by how independent each check is. These are counts, not a probability of correctness.

Evidence coverage and decision status are separate axes: an agent reporting 'done' never adds a passing check, and a human review or approval never counts as machine verification.

Dimension	Summary	Source
Task	Plan the retry policy; Write the retry tests · project payments-svc	client_log, mcp
Actors	claude-code · claude-opus-4-8	client_log
Actions	edit×5 execute×6 read×12 search×3 · touched 2 files · ran 2 commands	hook, mcp
Cost	$27.40 · pricing_table	client_log
Weekly plan	still learning — not enough 7-day history yet
Evidence	1 checks · 1 passed · 0 failed	mcp
Outcome	verified · asserted by machine	mcp

What ran

• Tools: Read×12, Bash×6, Edit×5, Grep×3
• Files touched: src/payments/client.py, tests/test_client.py
• Commands run: pytest tests/test_client.py -q, ruff check src/

Gaps (1) — what could not be proven

• evidence — 1 completed step has no linked passing check.

Provenance

• client_log — Observed in the agent's own local session / usage log.
• hook — Captured by an agentacct client hook (tool categories, mechanical checks).
• mcp — Recorded by the agent through agentacct's MCP tools (sections, files, checks).

Codex

Work Receipt — Add retry-with-backoff to the payments HTTP client

task_cx_retry

• Decision status — VERIFIED · asserted by machine Recorded machine evidence verifies the latest outcome.
• Evidence coverage — 1/2 self-checked · 1 unchecked Counts are passing checks over checkable steps, split by how independent each check is. These are counts, not a probability of correctness.

Evidence coverage and decision status are separate axes: an agent reporting 'done' never adds a passing check, and a human review or approval never counts as machine verification.

Dimension	Summary	Source
Task	Add backoff to the HTTP client; Run the payment tests · project payments-svc	client_log, mcp
Actors	codex · gpt-5.6-sol	client_log
Actions	edit×3 execute×4 read×9 search×2 · touched 2 files · ran 1 command	mcp, transcript_scan
Cost	$3.10 · pricing_table	client_log
Weekly plan	still learning — not enough 7-day history yet
Evidence	1 checks · 1 passed · 0 failed	mcp
Outcome	verified · asserted by machine	mcp

What ran

• Tools: read_file×9, exec_command×4, apply_patch×3, grep×2
• Files touched: src/payments/client.py, tests/test_client.py
• Commands run: pytest tests/test_client.py -q

Gaps (1) — what could not be proven

• evidence — 1 completed step has no linked passing check.

Provenance

• client_log — Observed in the agent's own local session / usage log.
• mcp — Recorded by the agent through agentacct's MCP tools (sections, files, checks).
• transcript_scan — Derived by agentacct from the client's own transcript / session store on disk (no live hook).

Reading the difference

Both agents finished the task, both recorded a passing test, and both receipts read verified with self-checked evidence — the receipt gives neither a stronger badge than it earned. What differs is how the work was observed and what it cost:

	Claude Code	Codex
Actions provenance	hook — a live client hook observed the tool categories	transcript_scan — read back from Codex's own on-disk session store (no hook fires)
Evidence tier	self-checked (agent-recorded via MCP)	self-checked (agent-recorded via MCP)
Cost (estimate)	$27.40 — an Opus-class model	$3.10 — a smaller model

Both checks are self-checked: the agent recorded the passing run itself. To raise a check to independently-checked, a client hook (or CI) has to observe it — the coverage matrix shows which agents support that. The point of the receipt is that it says self-checked here, instead of painting an agent-recorded pass the same green as an independently-verified one.

Note the two axes staying separate: the decision reads verified because the latest check passes and postdates the newest work, while the evidence coverage (1/2) shows not every step is individually checked. A clean decision word never hides partial coverage.

What this does not tell you (the honest unknowns):

• The costs are pricing-table estimates, not provider invoices.
• Self-checked does not mean wrong — it means no independent check observed the run.
• Neither receipt judges code quality or whether the retry design is right; they record what ran, what passed, and what it cost.

See also When an agent says done for how the timeline exposes a re-run and an outdated check, and the coverage matrix for what each agent's lanes can and cannot prove.