bench.yair-tech.com
← all runs

Run 002 · subagent orchestration

Claude Code conducting Grok 4.5 subagents. On tape, with the bill.

One Claude Code session gets one prompt: spawn two Grok 4.5 workers in parallel, have each build a browser game, verify the results. The workers run headless in their own isolated Linux container. Below: the uncut TUI recording (4× speed), the token meters on both sides, and what the same run would bill on metered APIs.

11 Jul 2026 · one take · 102 seconds wall clock end-to-end

The recording

The real session, sped up 4×: prompt in, both workers spawn in parallel (visible in the agent list at the bottom), files land, the conductor verifies and reports. Payoff: the two games it shipped.

The bill

Token counts as recorded. Both sides ran on flat-rate subscriptions (Claude Max, SuperGrok), so the actual marginal cost of this run was $0 — the metered column prices identical usage at each vendor's published API rates.

RoleWall clockTokensAPI-equivalentActual cost
Claude Fable 5 — conductor
Claude Code TUI · plans, spawns, verifies, reports
1:42 6.8k out · 662k cache read · 227k cache write1 $3.84 $0 · Claude Max
Grok 4.5 — builder ×2
grok CLI headless · isolated container · snake 0:24, breakout 0:36
0:36 (parallel) ~6k in · ~9k out2 $0.07 $0 · SuperGrok

1 Metered at Anthropic's published rates: $10/M input, $12.50/M 5-min cache writes, $1/M cache reads, $50/M output. 93% of the conductor's input was cache reads. The two thin Haiku relay wrappers that shell out to the container add roughly 56k mostly-cached tokens (<$0.05) on top. 2 The grok CLI reports no token usage headless — estimated from the actual session transcripts (chars÷4), priced at xAI's published $2/M in, $6/M out.

Method

  1. Claude Code (Fable 5) runs interactively in a 152×42 terminal, recorded with asciinema — the video is the real TUI, not a re-enactment.
  2. The conductor gets a single prompt asking for two parallel grok-worker subagents, one game each, then verification and a report.
  3. Each worker is a thin relay that dispatches the brief to a headless grok -m grok-4.5 inside an isolated Incus container.
  4. Games are published untouched. Token counts come from the Claude session log and the grok session transcripts; nothing is extrapolated beyond the noted estimate.