Nine Days of Claude and Codex Logs: Similar Session Counts, Different Token Accounting.

I didn’t choose to run two coding agents side by side. A subscription cutoff forced it. The old one was getting cancelled on a fixed date, and I needed the new one working before that date arrived — nine days with both, no choice involved.

Nine days in, someone asked me which one felt more efficient. I didn’t trust my gut on that, so I didn’t answer from memory. Both tools write every session to a local JSONL log. I read the logs instead.

The method, so the numbers mean something

I used ccusage, a log parser that reads Claude Code’s and Codex CLI’s local session files. It tallies tokens per day from both. I scoped it to September 30 through October 8 — the window both tools were actually active on my machine, not the two weeks I’d have guessed from memory.

Counts were close: 57 Claude sessions, 53 Codex sessions. That doesn’t establish equal tasks, complexity, or work completed. Session counts alone don’t make workloads comparable.

They didn’t.

The gap I didn’t expect

Metric (9 days) Claude Codex
Sessions 57 53
Reported input field; cross-tool definition unverified 5,024 19,220,514
Output tokens 2,092,063 2,090,128
Cache-read tokens 789,033,545 540,297,344
Reasoning tokens (visible) not reported 735,626

Reported output totals were both near 2.09 million. Input-field totals were 5,024 and 19,220,514. I haven’t verified that they exclude cached input in the same way. They aren’t a validated fresh-input comparison.

The arithmetic ratio is about 3,826. That doesn’t establish a difference of that size in new information, efficiency, or spending. Field definitions need to match first.

What I think explains it — labeled as a guess

Both reports contain large cache-read totals: 789 million and 540 million. I didn’t investigate the accounting internals. I can’t infer opposite caching behavior from the smaller input fields; zero can reflect reporting as well as activity.

That looks like a difference in how each tool’s log format accounts for repeated context. Not necessarily a difference in how much actual new information went in. I didn’t measure the underlying behavior directly, so I’m reporting what the log says, not what caused it.

Don’t read the cost column as money spent

Both ccusage reports include a cost estimate — $685.80 for Claude’s nine days, $553.05 for Codex’s. I almost dropped those numbers in without a caveat, and I’d have been misleading you.

Both of these are flat-rate subscriptions. I pay the same monthly price whether a session uses ten tokens or ten million. The cost column is ccusage estimating what the same token counts would cost on pay-per-token pricing — a comparison yardstick, not a bill I received. I don’t pay per-token for either tool, by design. If you take one number from this section, take the token counts, not the dollar figure.

Codex went quiet, and that’s data too

My Codex log stops on October 6. Not because I stopped using it. The CLI returned an error I didn’t choose: usage_limit_exceeded, with a message telling me to try again on October 10. Seven active days, then a hard wall, mid-week, no warning line before it hit.

Claude’s report covered the nine-day window without that recorded interruption. Codex returned a usage-limit error in this account’s October window. That’s historical account evidence, not a current outage report or a general plan-limit comparison.

The pattern that mattered more than tokens

Numbers aside, here’s what actually changed how I work: I stopped trusting either tool’s own “done” message without checking.

Twice this window, a task report said something was finished and it wasn’t fully true. Once, my own writing subagent reported it hadn’t published a draft when it had. I caught it by checking the live site instead of the agent’s summary. Separately, a fix report from the other tool described a UI feature as implemented. The build passed and the logic checked out, but the actual click-through on a real device was still unverified — a gap the report didn’t make obvious at a glance.

Neither case was a lie exactly. Both were a report conflating “the code is in place” with “I confirmed this works.” I don’t think that’s a defect specific to one tool. It’s a habit I now apply to anything either agent tells me is done: implemented, deployed, and verified-by-a-human are three different claims. I check which one I’m actually holding before I believe it.

What I’d tell myself before this started

I expected this comparison to produce a winner. It produced a session-count tie, an output-token tie, an input-token gap I can’t fully explain yet, and one hard usage wall I didn’t see coming. The more useful result wasn’t which tool “won.” It was building the reflex to read my own local logs instead of trusting how a week felt in hindsight.

Related: the GPT backend work behind this comparison is in this post.

Editorial correction, November 4, 2026: Removed the unverified fresh-input multiplier and workload equivalence, and dated the usage-limit observation. Original measurements are unchanged.

Similar Posts