Epoch AI puts agent costs at about $16–18 an hour for Codex, $24–50 for Claude Code

Epoch AI's agent-capacity report includes measured API-equivalent costs per hour of continuous agent work across Codex and Claude Code models.

Epoch AI's new report (October 2, 2026) estimates that AI chips shipped through 2027 could run tens to hundreds of millions of concurrent frontier-model agents. For builders, the more useful number is a budgeting benchmark. Epoch says that in its analysis of agent traces, Codex workloads averaged roughly $16–18 per hour of continuous agent activity, compared with $24–50 for Claude Code.

The figures come from TraceLab, a public dataset of logged Codex and Claude Code sessions. Epoch removed time spent waiting on humans, capped other idle gaps and priced tokens at API rates. Under its retained-cache and five-minute gap-cap assumptions, it reports these pooled rates:

Epoch then uses $30/hour as a round reference point for its capacity math.

Some caveats apply. These are API-equivalent costs, not subscription bills. SemiAnalysis argues that a plan's value depends on the plan, the model and the workload together, because subscription credit ratios can differ sharply from API price ratios. No source we found breaks the gap down into token efficiency, per-token price and harness behavior such as caching or subagents. We also found no independent test of Epoch's hourly figures.

A per-task view tells a similar but narrower story. In a repository-repair benchmark reported by Quasa, mean cost per attempted task at July 2026 API rates was $3.12 for Claude Code with Opus 4.8 and $2.34 for Codex CLI with GPT-5.5. The success rates were too close to rank reliably.

If you run long agent sessions on metered keys, divide your bill by active agent-hours and compare it against Epoch's range. Use Anthropic's pricing page to check the token and cache-write rates that go into your own number.

Sources