The anatomy of a 1M-token bill: $0.10, itemized to the token

Vendors love quoting per-token prices. Almost none will show you the actual bill a real workload produces. So here is ours: a full agentic coding session — the kind where you point an agent at a repo and let it work — with the complete itemized receipt, line by line, at the published rates.

Every number below is USD and comes from the live rate card at GET https://api.runaii.cloud/v1/models. You can reproduce this bill yourself, request by request, from your own console ledger.

The session

An agent working a medium-sized repo for an afternoon:

• 900,000 prompt tokens processed across ~180 requests — the system prompt, file contents, and tool outputs re-sent with each turn (this is what agents do; it's also exactly what prompt caching is for)
• 100,000 completion tokens written — diffs, file edits, summaries, tool calls
• 1M-token context on glm-5.3-flash — the whole session fits without compaction

The receipt

Line item	Tokens	Rate	Cost
Prompt — cache miss	200,000	$0.15 / 1M	$0.03000
Prompt — cache hit	700,000	$0.03 / 1M	$0.02100
Completion	100,000	$0.50 / 1M	$0.05000
Total	1,000,000		$0.10100

That's the whole bill. Ten cents for a million tokens of agentic work.

Where each fraction goes

The cache hit line is the story. 700,000 of the 900,000 prompt tokens were re-sends of prefixes the model had already seen — the system prompt, the file tree, tool outputs from earlier turns. At the published rates those billed at the cache rate, 5× cheaper than a miss:

Bucket	Rate (glm-5.3-flash)
Cache miss (first time a prefix is seen)	$0.15 / 1M tokens
Cache hit (same prefix seen again)	$0.03 / 1M tokens — 5× cheaper

Without caching, the same session costs:

0.900 × $0.15  = $0.135  (prompt, all miss)
0.100 × $0.50  = $0.050  (completion)
                = $0.185 total

With caching: $0.101. The split is automatic and reported per request in usage.prompt_tokens_details.cached_tokens — you can see exactly how many tokens hit the cache on every request.

What the same session costs on the frontier sibling

Same session, glm-5.3 (the frontier agentic model, same 1M context):

Line item	Tokens	Rate	Cost
Prompt — cache miss	200,000	$1.40 / 1M	$0.28000
Prompt — cache hit	700,000	$0.26 / 1M	$0.18200
Completion	100,000	$4.40 / 1M	$0.44000
Total			$0.90200

Both rows are the same bill shape; pick the model per job. Flash for high-volume pipelines, frontier for the hard steps. The rate card carries both.

What happens when a stream dies mid-flight

You pay for delivered tokens only. Every request moves through three money events:

1. Authorize — the estimated worst case is reserved from your balance, atomically. Two concurrent requests cannot double-spend the same credits.
2. Settle — when the stream finishes, the reservation is settled against actual usage: prompt tokens at the full rate, cached tokens at the cache rate, completion tokens at the output rate.
3. Reverse — if the request died mid-stream, the whole reservation is refunded. You never pay for a request that never delivered.

In this session, 4 of the 184 requests hit provider errors. All 4 show up as explicit refund events in the ledger — $0.00041 reversed in full, visible line items, not silent adjustments.

Verify this receipt yourself

• usage.cost in every response — per-request truth
• Console → Usage — per-model spend, latency percentiles, daily charts
• Console → Billing — the full ledger: every charge, top-up, and refund, with the request id that caused it

Re-run this session shape yourself and compare the receipt to this post. If the ledger doesn't match the post, the receipt wins and the post gets corrected.

---

This post is part of the runaii.cloud transparency series. The rates above are the live output of GET /api/v1/models as of 2026-10-01 — see the price comparison post for the full rate card, and the billing engine post for the reserve → settle → refund internals.