Token Spend & Value Analysis review

Key Observation: Platform Shift

Claude Code dominated Jan–Mar 2026 (sole primary tool). Codex adoption exploded in April — going from 176 prompts in March to 3,851 in April. By May, Codex surpassed Claude in raw prompt volume. This coincides with Codex becoming production-ready and your shift to parallel autonomous workflows.

Claude Code usage dropped ~37% from April to May, suggesting active migration of certain workflow types to Codex rather than net-new capacity.

Trend: Sessions Got Shorter, Tool Use Got Denser

Messages per session dropped from 300 (Feb) to 118 (Apr) while tools per message increased from 0.15 to 0.50. This means sessions became more focused and tool-heavy — less conversational, more execution-oriented. Likely driven by skill adoption and shipping conventions.

Claude Code: Interactive Development Hub

Claude Code is your primary interactive development tool. The heaviest spend is on:

• bismarck-v0.4 (1,228 prompts) — game project, heavy /ship usage
• loadoutworks.com (962 prompts) — web project, active shipping
• metternich-engine (906 prompts) — game engine, active development
• lexcorp-war-room (694 prompts) — active project with /investigate usage
• /ship is 79% of all skill calls — the shipping automation is by far the most-used workflow

Codex: Autonomous Batch Processor

Codex handles autonomous, parallelizable work. Top spend:

• mobile-ideas (106 sessions, 321M tokens) — largest Codex consumer by far
• agentic-skills (47 sessions, 282M tokens) — skill development and meta-work
• pitwall-monorepo (68 sessions, 190M tokens) — monorepo operations
• Average cost per session: $1.10 (very cheap)
• Average cost per prompt: $0.05 (negligible)

Value Verdict

Claude Code (~$6.9K) delivers value through interactive development, debugging, and shipping automation. The /ship skill alone (1,972 calls) automates commit+push+docs workflows that would take 5–10 min each manually — conservatively saving ~165 hours of shipping ceremony. At a $100/hr developer rate, that's ~$16,500 in avoided manual work from /ship alone.

Codex (~$360) delivers exceptional token efficiency. At $1.10/session and $0.05/prompt, it's a bargain for autonomous batch work. The high cache rate (94.4%) means you're barely paying for repeated context reads. The value proposition is clear: Codex handles the long-tail of parallel, autonomous tasks that don't need real-time interaction.

Combined, the tools complement each other well — Claude for interactive/creative/debugging work, Codex for autonomous/batch/parallel work. The shift from Claude-heavy (Jan–Mar) to Codex-augmented (Apr–Jun) suggests you've found a productive division of labor.

1. Codex Input/Output Ratio (215:1)

Observation: Codex reads 215 tokens for every 1 it writes. The costliest sessions hit 400–600:1.

Implication: Even with 94.4% caching, the sheer volume (1.05B input tokens) costs $338 in input alone. If the non-cached fraction could be reduced from 5.6% to 3%, monthly Codex spend drops by ~40%.

Action: For agentic-skills (the project with the highest tokens-per-prompt at 1,369K), consider whether Codex sessions are re-reading full repo context unnecessarily. Scoping base instructions or using smaller working sets could cut input tokens significantly.

2. Claude Code Feb Spike ($2,310)

Observation: February was the costliest month — 220,660 messages, 300 messages/session (longest average sessions).

Implication: Long sessions accumulate context that pushes up per-message cost. The Feb spike is likely from early-stage project development with lots of back-and-forth.

Action: The trend already improved — Apr dropped to 118 msgs/session. The natural shift to shorter, tool-heavy sessions is the right trajectory.

3. Skill Coverage Gap

Observation: /ship (79%) dominates skill usage. /code-review and /expert-review together are only 23 calls — surprisingly low for the volume of code being produced.

Action: Consider whether more systematic code review before /ship would catch issues earlier. The /investigate count (120) suggests bugs are being found after the fact.