Codex analytics for a team, by repository

Codex analytics for a team lives in OpenAI's Admin Console, on ChatGPT Enterprise: active users, credits and tokens, threads and turns, a user leaderboard and model usage, 120 days back, with an API behind it. A Business workspace gets Workspace analytics, a member table and a link into Codex analytics; one person gets /status and the usage dashboard. None of them has a row per repository, and none shows Codex beside Claude Code. Caprock does both, from the session files on each machine.

Codex analytics for ChatGPT Enterprise: what an admin sees

OpenAI's Codex pricing page lists, under "Admin, security, and analytics", the analytics dashboard, the Analytics API, and the Compliance API with audit logs for Enterprise and Education only. Read on 30 October 2026, what each holds:

• The dashboard. The Global Admin Console added Codex analytics on 21 May 2026 "with active users, credits and tokens, threads and turns, user leaderboards, plugin usage, accepted lines of code, model usage". Since 18 June 2026 it exports "credits used by Codex and ChatGPT, leaderboards for users and agents, and Codex and ChatGPT usage broken down by feature"; since 16 July 2026 it keeps 120 days; since 23 July 2026 leaderboards have a Groups tab with "group-level breakdowns by product, metered item, and model", and "estimated dollar values may appear next to credits in Analytics, Billing, usage-limit editors, the Cost API, and the Codex analytics API. Estimates use the selected workspace’s overage rate and are for planning." The console article gives Codex usage analytics a window of "Up to 120 days", refreshed "Approximately every 1 to 6 hours", and adds: "Codex product access does not automatically include analytics."
• Usage Insights groups the same activity into use cases and tasks, with messages, share of credits and active users per category, model, reasoning and speed breakdowns, and a Code review tab: PRs reviewed, issues found, reactions. OpenAI's own caveats: "A message count doesn’t tell you how much work the team finished." and "A reviewed pull request isn’t necessarily merged or deployed."
• The Codex analytics API. "Codex Enterprise Analytics is available to Enterprise workspaces with Codex enabled." An owner creates a workspace-scoped Admin key under Credentials > Admin keys with codex.enterprise.analytics.read; the reference lists four calls: Get Codex Turn Insights, Get daily Codex usage, Get daily Codex code review metrics, Get daily Codex code review user responses. A second endpoint, under enterprise.analytics.usage.read, "reports daily ChatGPT, Codex, and Work usage."
• The Compliance API is for records, not reporting. "The Compliance Platform is available to ChatGPT Enterprise and Edu workspaces." Codex "usage, including local clients such as the CLI and IDE extension as well as web or cloud-delegated usage, is available in the Compliance API." OpenAI's docs say of it: "It’s not a productivity dashboard. Don’t use it to infer code quality or individual performance."
• Spend controls. "Usage limits control each person's usage through workspace defaults, group defaults, and individual overrides. The amounts are in credits or USD, depending on your billing model."

Every dimension on that list is a person, a group, a model or a feature. The help centre's export schema says the same: the Workspace analytics usage exports are a User report, a GPT report and a Project report, the project being a ChatGPT project; the fourth export is an Impact survey report.

Codex analytics on ChatGPT Business

The Business release notes of 16 April 2026 describe Workspace analytics for Business: "View headline metrics such as active users, total messages sent, and total credits spent across the selected time range." A member table: "See usage by workspace member, including seat type, credits spent, and messages sent." And a link out: "Jump directly to Codex analytics for lightweight insights into credits usage and developer activity." The Codex pricing page, read the same day, marks the analytics dashboard and both APIs for Enterprise and Education only, and gives Business, in that section, "SAML SSO, MFA, and workspace user management", managed configuration and the no-training default, with no analytics row. Read together: Business sees its members and their credits; the API, the 120-day console history and the Compliance API are Enterprise.

Spend controls on Business are "monthly credit usage limits by seat type and per-user overrides", and "By default, all seats and users have no limits specified."

Codex Business and Enterprise pricing, for seats

What one person sees

/status in a Codex CLI session, and the usage dashboard at chatgpt.com/codex/settings/usage, show one account's five-hour and weekly windows. On an eligible Enterprise workspace, ChatGPT Desktop's Settings > Usage & billing adds a monthly credit meter, credit and token history grouped by Surface, Model, Reasoning or Speed, and, per chat, "Lifetime credits used by the Work or Codex chat" and "An estimated cost, when available". The plans, the dashboard and what a session costs: Codex usage and cost. The windows by plan: Codex usage limits.

What none of them shows

• By repository. No page read on 30 October 2026 names a repository as a dimension of Codex usage or spend. The console knows users, groups, models and features; the exports know users, GPTs and ChatGPT projects.
• By pull request. The pull requests in Usage Insights are the ones Codex reviewed, counted, with no cost beside them and no link to the sessions that wrote them.
• Beside Claude Code. OpenAI's analytics cover OpenAI's agent. A team running Codex and Claude Code has two consoles and no row that adds them up.
• A session on an API key. Of the workspace's spend controls, OpenAI's docs say: "They don’t cover all Codex usage or govern OpenAI API Platform billing." A developer who switched the CLI to a key is on the Platform bill, not in the workspace.

Codex and Claude Code in one table

Caprock records every session on a machine, whichever agent ran it and wherever it was started, and prices each turn at the vendor's list rate. Thirty days on my Mac, read from its database on 28 October 2026:

The same 2,363 sessions, by who started them:

626

sessions started from Caprock, $6,619

1,737

started in a terminal or elsewhere, $790

A cockpit lists the sessions it started; the 1,737 others would not be on it. A recorder has both. In Caprock each session sits in its repository's row, the repository's cost is divided by its merged pull requests, and Codex and Claude Code add up in the same row.

How Caprock reads Codex

Codex writes one append-only JSONL transcript per session under $CODEX_HOME/sessions/ (~/.codex by default) and moves archived threads to archived_sessions/. Caprock tails those files, compressed .jsonl.zst included, from the CLI and the desktop app alike: no hook, nothing written into Codex's config, no process signalled, nothing leaving the machine. From each file:

• session_meta gives the session id, the working directory and the client. The directory is how a session lands in a repository's row.
• token_count gives input, cached input, output and reasoning tokens per turn, and the plan-limit windows. Deltas come from the cumulative field, which reproduces Codex's own total exactly; cached tokens are taken out of the input count so nothing is priced twice.
• The price comes from Caprock's dated table of OpenAI's list rates. Codex itself never writes a cost.

Where the figure stops:

• gpt-6 cache writes. OpenAI's API price for the gpt-6 family bills cache writes above the input rate (1.25× input for a 30-minute cache, 2.5× for 12 hours, read 30 October 2026), but Codex reports no cache-write tokens, so there is nothing to apply the rate to: for those turns the list-price figure can be low by up to a quarter of that unreported share at the rate Caprock's table applies. On a ChatGPT plan the point does not arise: "Codex credit billing has no separate cache-write charge. API-key usage follows API pricing."
• A total with no breakdown is priced as input, an upper bound, and the turn is marked so.
• Above 272K input tokens the API's long-context surcharge is not modelled; no observed turn reached it.
• Codex's background reviewer is counted in tokens on its own line with no price, because OpenAI publishes none.

Across the team

Caprock for Teams is the ledger across agents and machines: every Codex and Claude Code session on every developer's machine, priced the same way, summed per repository, pull request and release on a server you run, with no field for a prompt, a reply or a command's output. It sits beside OpenAI's console, which keeps the credits and the people, and beside Anthropic's. Buying Caprock is one flat price for the team.