What one piece costs — the current process, step by step, metered on the gifts piece measured 2026-09-19 · recipe r33 · 12.7¢ · 6½ minutes
One article written under the current recipe with a meter on every step: the spend was read before and after each desk call, so each box below carries what that step actually cost on this run. Prices are DeepSeek V4.1 Flash off-peak, Gemini 3.8 Flash for the defect reader, and the search engine's fee per query. Cached prompt prefixes make the persona's own steps nearly free; the searches are not cached and are most of the bill.
the answer in one line12.7¢ a piece, 6½ minutes. Half of it is looking things up (6.1¢, of which about 5¢ is ten search-engine queries at half a cent each), an eighth is the witness desk (1.6¢, three more searches), an eighth is the defect reader (1.75¢, three Gemini reads), and the writing itself — draft, reread, two mends — is 2.5¢. Sixteen pieces a morning would be about $2.
The words. A step is one box below; a step may be several calls. Cached means DeepSeek serves the unchanged front of the prompt (the persona's profile, the commission) at a tenth of the price, which is why the persona's short steps cost a fraction of a cent. A search is one query to the web search engine; the pages it returns are fetched by us for nothing. The $0.005 metered here is Brave's price: on the day of the measurement the Serper account had run out of credits, so every English query fell through to Brave (five times Serper's $0.001), and the brew booked it under Serper's name until the same-day fix that books each query under the engine that answered. With Serper topped up, the search line is about a fifth of what is shown below. The reader is a model that is not the writer.
The flow, with the cost on each step
the reportingthe desk's notebook, already on file for this commission0¢ here
↓
1 · the reactionwhat struck the persona, what it wants to know · 1 call, cached prefix0.07¢ · 6 s
↓
1b · the life recalleda reader scores the persona's 93 stored episodes and positions against the news · 1 call0.23¢ · 8 s
↓
2 · its questions looked up10 searches (6 its own + 4 canned) · 12 pages read · 12 notes calls · 1 merge6.11¢ · 122 s — of which ~5¢ is search fees
↓
3 · the pointwhat it holds, having looked · 1 call, thinking on0.09¢ · 5 s
↓
3a · the point's two testsabsence by code · mechanism by a reader0.01¢ · 1 s
↓
3c · the witness desk3 searches · 8 pages · quotations pulled and checked verbatim1.59¢ · 24 s — of which 1.5¢ is search fees
↓
4 · its own outline1 call0.05¢ · 2 s
↓
5 · the draft, written whole1 call, thinking on, ~900 words · + the title0.51¢
↓
8 · the defect reader marksGemini 3.8 Flash reads the draft with the taxonomy · 1 read per round + 1 after1.75¢ · 44 s for three reads
↓
8 · the mendthe writer rewrites whole with the marks in view · 2 rounds0.96¢ · 61 s
↓
6 · the rereadthe writer reads its own piece and writes it again · 1 call, thinking on1.00¢ · 68 s
↓
7 · the checktwo stages on DeepSeek Flash: list the checkable statements, verdict them · then the writer's correction0.21¢ + 0.15¢ · 14 s
↓
5 · the prior recordedthe point kept under the persona's name for its next piece · no call0¢
the persona writes or decidesa reader that is not the writerthe desk searches and reads pagesa check
The same bill, by kind of cost
| kind | ¢ | share | what it buys |
| search-engine fees | 6.5 | 51% | 13 queries at $0.005: ten for the persona's questions and the canned four, three for the witness |
| the desk's reading (DeepSeek, no cache) | 1.4 | 11% | notes on 12 pages, the answers merged, the witness quotations pulled, the recall |
| the defect reader (Gemini) | 1.75 | 14% | three reads of a 900-word piece with the taxonomy |
| the writing (DeepSeek, cached prefix) | 2.5 | 20% | draft, two mends, the reread — the profile is served from cache each time |
| the persona's short steps | 0.25 | 2% | reaction, point, outline, tests, correction |
| one piece | 12.7 | 100% | 6½ minutes, 899 words |
Where the money could move
- The searches are half the bill only because Serper was empty. At Serper's $0.001 a query the same 13 searches are 1.3¢, and the piece is about 7½¢. Topping up Serper is the cheapest saving on this page. Then: Ten questions went out; six came back answered. Searching only the persona's own questions and the two canned ones it never asks (elsewhere, critics) would save about 1¢ with little loss; the witness desk's three searches are what put a real person in the piece and should stay.
- Time, not money, is the constraint for a morning edition. 6½ minutes a piece, of which two minutes is the look-up because searches run one a second by design (the engine refused sixty at once). Sixteen pieces in parallel still finish inside twenty minutes; the searches' one-a-second door is the only serial part, at about 3½ minutes for 208 queries.
- The defect reader is the one place a stronger model would be worth trying by price: three Gemini reads are 1.75¢; a Claude Haiku read would be about 1¢ each, and the owner's read found six false positives among the current reader's marks.
- Compared with a week ago: the old process (three drafts, code picks, code repairs) cost about 3¢ a piece and read as a machine every time. The new one is four times the money for a piece the owner reads through. The whole pipeline's daily bill, at sixteen pieces, is about $2.