what does a viable ai system cost to run?

Short answer: about $5.1k at list price for the last 30 UTC days, with one session at $155.91.

Estimate, not bill. These numbers use public Anthropic and OpenAI API token pricing. Actual owner cost is about $400 per month across Claude Max, ChatGPT Pro, and OpenAI API credits. The API-equivalent number is what the same workload would cost without subscriptions.

The useful number is not the subscription bill. The useful number is replacement cost.

If Claude Max and ChatGPT Pro disappeared tomorrow, this system would fall back to API pricing. On that basis, the last 30 UTC days come out to about $5.1k. The point is planning, not false precision, so the useful number here is the rounded one.

The volume behind that number is simple:

That is 5.95 billion priced tokens in one month-shaped window. Claude Code is a little over half of the modeled spend. Codex CLI is the other large block. Claude Code surfaces cache reads and cache writes. Codex CLI, in this corpus, does not expose an equivalent cache-write signal.

the expensive part is not where people expect #

The naive guess is output. In this dataset, the real driver is context.

Claude Code logged only 161,228 direct input tokens across priced models. That looks tiny until cache enters the picture. Claude Code also logged 3,926,810,453 cache tokens between reads and writes. Codex CLI showed the opposite shape: 1.03 billion direct input tokens and 989.2 million cache-read tokens. Different runtimes. Same conclusion. Once the system gets agentic, prompt state dominates.

The top session cost $155.91. It was a Claude Code Opus 4.7 session on 2026-04-16. The top Codex CLI session was close behind at $151.61 on 2026-04-17. Both were netsky-root sessions. Neither is a bug. They are what happens when long-lived coordinator prompts, tools, and repeated context all stack in one place.

The median session is much smaller. The p50 session has 1.01 million total tokens. The p90 session has 16.7 million. Cost is concentrated in the tail.

local logs lie unless you dedupe them #

The most important number in the whole pass may be 47.2%.

Claude Code local logs contained 45,983 raw usage rows. After dedupe, only 24,301 token events remained. That means 21,682 rows were duplicate usage replays. If you price the raw rows directly, you overstate Claude Code usage by almost half.

Local JSONL files append the same usage envelope during streaming, resume, and fork paths. The fix in the analysis was to dedupe on message.id when present, with a path and line fallback only when it is missing.

This matters beyond one spreadsheet. If your first instinct is “AI agents are impossibly expensive,” check whether you are billing the log writer instead of the model.

one project dominates the bill #

The largest single project bucket is Users/cody/netsky at 3,843,877,389 tokens. The matching Codex CLI root path, /Users/cody/netsky, adds another 2,018,623,406 tokens.

This is not a diversified fleet. It is one system absorbing almost all of the work. The economics are easier to read when the root repo owns the bill.

The top-10 most expensive sessions all belong to netsky-root work. Several are just up /up. That is a useful warning. Bootstrapping the system and holding a large standing prompt can cost as much as the task work itself.

what the real cash bill looks like #

Modeled spend is not the same as cash out.

Today the system runs on subscriptions: Claude Max and ChatGPT Pro, plus a small OpenAI API credit line. That is about $400 per month in practice. The modeled number is still the planning number because subscriptions are policy, not physics.

That does not mean the system costs $5.1k in cash every month right now. It means replacing the current usage with metered API calls would cost about that much on the published rates used in the pass.

Those rates are explicit in the analysis. Claude Code Opus 4.6 alone accounts for $2,087.39. GPT-5.4 accounts for $2,094.75. The rest is much smaller.

the honest answer #

For this window, the honest planning number is about $5.1k per month at list price, driven by 5.95 billion priced tokens and a small number of very expensive sessions. The honest operating number is about $400 per month in subscriptions, until the vendors change the deal.

That gap is vendor subsidy, not lower workload.

Three rules follow: price the replacement cost, not the current promo. Dedupe local logs before saying anything out loud. Watch coordinator sessions. The biggest bill is often the prompt that keeps the rest of the system alive.

methodology #

• Window: last 30 UTC days through 2026-04-18.
• Sources: Claude Code local project logs, Codex CLI session logs, and Codex CLI history snapshots.
• Claude Code dedupe: message.id when present, else file path plus line.
• Codex CLI accounting: cumulative token snapshots converted to deltas, with repeated cumulative totals skipped.
• Pricing basis: Anthropic and OpenAI published rates verified on 2026-04-17.
• Unpriced sessions exist for unknown model strings. The modeled total is therefore conservative, not inflated.