# OpenAI reading 0: OpenAI's own history, measured on our tasks

*The OpenAI lane of the cost-of-intelligence instrument, 2026-09-23 (UTC). Every table below
is generated by `tables.py` from the row files and the saved receipts in this directory; the
prose is the lane's. Tables regenerated 2026-09-23T18:18:15Z.*

## What this measures, for a reader who scrolled straight here

OpenAI still serves most of its own history on an operator's API key: `gpt-3.5-turbo` snapshots from
late 2023, `gpt-4-0613` (the June 2023 frontier), and every flagship since, through the
`gpt-6` line of September 2026. This reading asks those models the **same frozen questions**
the local bench asked the gemma4 and OpenJev models on benchbox's 3090s, through the **same
harness**, so the frontier's own curve can be drawn on our tasks instead of borrowed from
someone else's benchmark.

One task was read, from the Jev bench's frozen kit (series set S0, R-1):

- **task c, the six-way (108 items).** A line spoken at a long-table dinner; which of six
  guests said it. Chance is 16.7 %.

A second, **task d, the doorman planted set (48 items)**, was planned and is **HELD**: the set
is private (`doorman-planted-2026-09-13/harness/planted_set.py:5`, "never go anywhere
public"), and at 2026-09-23 ~17:45Z the lead held its egress to any API pending ruling ⧖-4
(PLAN-v2 §7 item 0). **No doorman item was sent to OpenAI**: the pilot, the probes and the
default-effort reads used six-way items or the one-line probe prompt only, and the guard in
`run_openai.py` (no `../DOORMAN-EGRESS-RULED` marker → print `HELD on ⧖-4`, exit 0) turned
phases D and E into no-ops before either began.

The window: pricing read 2026-09-23T15:40:55Z; probes 15:48Z–15:49Z; pilot 15:49Z–15:59Z;
full rows 16:00:42Z to 18:03:37Z (the six-way; phases D and E held, below). The Ollama Cloud and Anthropic readings sit in the
sibling directories; the local rows this is compared with are named below with their files.

## The headline

**The six-way, 108 items, one reading each, in release order.** Every figure below is
`correct / 108` from `rows/<model>.c.report.json`; chance is 18 of 108.

- **The 2023 frontier, `gpt-4-0613`, scored 88 of 108 (81.5 %, 95 % 73.1–87.7 %), for
  $0.9553, or $0.01086 per correct answer.** Its launch price and today's price are the
  same ($30 / $60 per million tokens), so that figure is also what it cost in June 2023.
- **The OpenAI flagships then climb and flatten.** `gpt-4-turbo` 92 (April 2024),
  `gpt-4o-2024-05-13` 88, `gpt-4.1` 97 (April 2025), `o3` 98, `gpt-5` 97, `gpt-5.2` 96 and
  `gpt-5.4` 98 (March 2026) sit on a plateau near 89–91 %. `gpt-5.5` reached 102 and
  `gpt-5.6-sol` 103 (95.4 %), the top of this reading. `gpt-6-astra` and `gpt-6-sol` both
  scored 101. The late steps are one or two items apart, well inside each other's
  intervals.
- **The 2023 bar now costs a few hundred times less.** Of the models at or above the
  anchor's 88, the cheapest per correct answer is `gpt-6-luna`: 96 of 108 for $0.0032 in
  total, or $0.000033 per correct answer. That is about 326 times cheaper than the anchor,
  three years and three months later. `gpt-5.6-luna` (94, $0.000069 per correct) and
  `gpt-5.4-mini` (90, $0.00030) follow.
- **The 3.5 line was never on this curve.** `gpt-3.5-turbo-1106` scored 34 and `-0125`
  scored 24, both months after `gpt-4-0613`'s 88. They were the cheap tier of their day,
  not the frontier, and on this task they are close to chance.
- **On our own cards:** benchbox's `gemma4:26b` generate arm, on one RTX 3090, scored 89 of
  108 (88 strict). That is the 2023 frontier's score on this task. OpenJev FP8 (two 3090s)
  scored 101, which is `gpt-6-astra`'s and `gpt-6-sol`'s score. Both comparator rows are
  named with their files in "Beside the local rows".

## How it was asked (the instrument, and what is verbatim)

- **The harness is the Jev bench's own generate path, imported and never edited**
  (`jev-2026-09-21/run.py`, and `run_addenda.py` for task d). The prompt builder, the
  instruction strings, the option letters and order, the letter parse (the first A–F
  character anywhere in the reply) and `summarise()` are that file's. Every report records
  the sha256 of `run.py`, `run_addenda.py`, both kit files and this lane's two scripts.
- **The OpenAI call** (`run_openai.py` + `oai.py`): Chat Completions, one call at a time,
  streamed with `stream_options.include_usage` (R-10c), 0.2 s between calls (0.35 s in the
  pilot), backoff on 429
  and 5xx. Temperature 0 and seed 20260921 wherever the model accepts them.
- **The token limit.** A model with no reasoning gets `max_tokens = 8`, which is `run.py`'s
  own `predict=8`; the local gemma4 generate arm had the same eight tokens. A reasoning model
  gets a 4,096-token `max_completion_tokens` budget (16,384 at its default effort), because
  its hidden reasoning bills against that budget and eight would truncate it into a
  non-answer. A reasoning model's visible reply is therefore not cut at eight tokens; the
  harness reads it the same way.
- **The effort.** Each model was probed (`receipts/profiles/`) for the lowest
  `reasoning_effort` it accepts, trying `none`, then `minimal`, then `low`, and every refusal
  is kept with the server's own sentence. R-11 adds two flagships read again at the server's
  default effort.
- **Scoring rules the harness imposes, stated so nobody has to find them.** A reply with no
  option letter is scored by `run.py` as option A (uniform probabilities, first index). That
  is how the local comparators were scored, so it is the primary figure here too; the
  **strict** column counts such a reply as wrong. A failed call (after backoff) and a
  reasoning model that runs out of budget before any visible answer are instrument
  artefacts: they are written as refused rows and kept out of every rate, never scored as
  wrong. Neither happened in this reading (the failed-calls and truncated columns).
- **Prices.** Today's prices are parsed from the saved pricing page
  (`receipts/pricing.md`, fetched from https://developers.openai.com/api/docs/pricing.md at
  2026-09-23T15:40:55Z), standard tier, short context. Every call's dollars are its own usage
  block times that row (`receipts/spend.jsonl`). A dated snapshot with no row of its own is
  priced from its alias row and says so.

## Prices: today's, and at launch (R-4)

"date" is the snapshot date the model ID names, or the API's own `created` date where the ID
names none (`receipts/models-list.json`). Launch prices come from OpenAI's launch posts, its
changelog, or the Wayback Machine's copy of its pricing page nearest after launch, each saved
under `receipts/launch-prices/` with the quoted line in `launch-prices.json`; all 33 quotes
were checked against their saved page text. The 2023–early-2024 rows were quoted per 1K
tokens and are shown ×1,000. Note that the API `created` dates for the 2026 undated IDs are
earlier than OpenAI's public release dates in its changelog (gpt-5.6: created 06-23,
released 07-09; gpt-6-astra: 08-27 vs 09-03; gpt-6-sol/luna: 09-14 vs 09-22).

| model (API id) | date | date from | today in $/1M | today out $/1M | today's page row | price verified | launch in $/1M | launch out $/1M | launch source | launch source date | launch source URL |
|---|---|---|---:|---:|---|---|---:|---:|---|---|---|
| gpt-4-0613 | 2023-06-13 | ID | $30.00 | $60.00 | gpt-4-0613 | yes | $30.00 | $60.00 | wayback-pricing | 2023-06-14T10:42:37Z | https://web.archive.org/web/20230614104237/https://openai.com/pricing |
| gpt-3.5-turbo-1106 | 2023-11-06 | ID | $1.00 | $2.00 | gpt-3.5-turbo-1106 | yes | $1.00 | $2.00 | wayback-pricing | 2023-11-07T00:22:12Z | https://web.archive.org/web/20231107002212/https://openai.com/pricing |
| gpt-3.5-turbo-0125 | 2024-01-25 | ID | $0.50 | $1.50 | gpt-3.5-turbo-0125 | yes | $0.50 | $1.50 | announcement | 2024-01-26T21:26:18Z | https://web.archive.org/web/20240126212618/https://openai.com/blog/new-embedding-models-and-api-updates |
| gpt-4-turbo-2024-04-09 | 2024-04-09 | ID | $10.00 | $30.00 | gpt-4-turbo-2024-04-09 | yes | $10.00 | $30.00 | wayback-pricing | 2024-04-10T02:33:50Z | https://web.archive.org/web/20240410023350/https://openai.com/pricing |
| gpt-4o-2024-05-13 | 2024-05-13 | ID | $5.00 | $15.00 | gpt-4o-2024-05-13 | yes | $5.00 | $15.00 | wayback-pricing | 2024-05-14T20:10:30Z | https://web.archive.org/web/20240514201030/https://openai.com/api/pricing |
| gpt-4o-mini-2024-07-18 | 2024-07-18 | ID | $0.15 | $0.60 | gpt-4o-mini | yes | $0.15 | $0.60 | announcement | 2024-07-19T20:08:26Z | https://web.archive.org/web/20240719200826/https://openai.com/index/gpt-4o-mini-advancing-cost-efficient-intelligence/ |
| gpt-4o-2024-08-06 | 2024-08-06 | ID | $2.50 | $10.00 | gpt-4o | yes | $2.50 | $10.00 | announcement | 2024-08-07T17:23:51Z | https://web.archive.org/web/20240807172351/https://openai.com/index/introducing-structured-outputs-in-the-api/ |
| gpt-4o-2024-11-20 | 2024-11-20 | ID | $2.50 | $10.00 | gpt-4o | yes | $2.50 | $10.00 | wayback-pricing | 2024-11-29T21:20:29Z | https://web.archive.org/web/20241129212029/https://openai.com/api/pricing |
| o1-2024-12-17 | 2024-12-17 | ID | $15.00 | $60.00 | o1 | yes | $15.00 | $60.00 | wayback-pricing | 2024-12-21T05:29:46Z | https://web.archive.org/web/20241221052946/https://openai.com/api/pricing |
| o3-mini-2025-01-31 | 2025-01-31 | ID | $1.10 | $4.40 | o3-mini | yes | $1.10 | $4.40 | wayback-pricing | 2025-02-01T14:59:21Z | https://web.archive.org/web/20250201145921/https://openai.com/api/pricing |
| gpt-4.1-2025-04-14 | 2025-04-14 | ID | $2.00 | $8.00 | gpt-4.1 | yes | $2.00 | $8.00 | announcement | 2025-04-15T18:27:20Z | https://web.archive.org/web/20250415182720/https://openai.com/index/gpt-4-1/ |
| gpt-4.1-mini-2025-04-14 | 2025-04-14 | ID | $0.40 | $1.60 | gpt-4.1-mini | yes | $0.40 | $1.60 | announcement | 2025-04-15T18:27:20Z | https://web.archive.org/web/20250415182720/https://openai.com/index/gpt-4-1/ |
| gpt-4.1-nano-2025-04-14 | 2025-04-14 | ID | $0.10 | $0.40 | gpt-4.1-nano | yes | $0.10 | $0.40 | announcement | 2025-04-15T18:27:20Z | https://web.archive.org/web/20250415182720/https://openai.com/index/gpt-4-1/ |
| o3-2025-04-16 | 2025-04-16 | ID | $2.00 | $8.00 | o3 | yes | $10.00 | $40.00 | wayback-pricing | 2025-04-17T00:34:54Z | https://web.archive.org/web/20250417003454/https://openai.com/api/pricing |
| o4-mini-2025-04-16 | 2025-04-16 | ID | $1.10 | $4.40 | o4-mini | yes | $1.10 | $4.40 | wayback-pricing | 2025-04-17T00:34:54Z | https://web.archive.org/web/20250417003454/https://openai.com/api/pricing |
| gpt-5-2025-08-07 | 2025-08-07 | ID | $1.25 | $10.00 | gpt-5 | yes | $1.25 | $10.00 | announcement | 2025-08-09T02:19:39Z | https://web.archive.org/web/20250809021939/https://openai.com/index/introducing-gpt-5-for-developers/ |
| gpt-5-mini-2025-08-07 | 2025-08-07 | ID | $0.25 | $2.00 | gpt-5-mini | yes | $0.25 | $2.00 | announcement | 2025-08-09T02:19:39Z | https://web.archive.org/web/20250809021939/https://openai.com/index/introducing-gpt-5-for-developers/ |
| gpt-5-nano-2025-08-07 | 2025-08-07 | ID | $0.05 | $0.40 | gpt-5-nano | yes | $0.05 | $0.40 | announcement | 2025-08-09T02:19:39Z | https://web.archive.org/web/20250809021939/https://openai.com/index/introducing-gpt-5-for-developers/ |
| gpt-5.1-2025-11-13 | 2025-11-13 | ID | $1.25 | $10.00 | gpt-5.1 | yes | $1.25 | $10.00 | wayback-pricing | 2025-11-18T17:00:45Z | https://web.archive.org/web/20251118170045/https://openai.com/api/pricing |
| gpt-5.2-2025-12-11 | 2025-12-11 | ID | $1.75 | $14.00 | gpt-5.2 | yes | $1.75 | $14.00 | announcement | 2025-12-13T00:51:22Z | https://web.archive.org/web/20251213005122/https://openai.com/index/introducing-gpt-5-2/ |
| gpt-5.3-chat-latest | 2026-02-27 | API created | $5.00 | $30.00 | gpt-5.5 (<272K context length) | NO — unverified: no row on the page; costed at the family's highest non-pro row (gpt-5.5) | $1.75 | $14.00 | wayback-pricing | 2026-03-04T03:03:14Z | https://web.archive.org/web/20260304030314/https://developers.openai.com/api/docs/pricing |
| gpt-5.4-2026-03-05 | 2026-03-05 | ID | $2.50 | $15.00 | gpt-5.4 (<272K context length) | yes | $2.50 | $15.00 | wayback-pricing | 2026-03-06T01:09:35Z | https://web.archive.org/web/20260306010935/https://developers.openai.com/api/docs/pricing |
| gpt-5.4-mini-2026-03-17 | 2026-03-17 | ID | $0.75 | $4.50 | gpt-5.4-mini | yes | $0.75 | $4.50 | wayback-pricing | 2026-03-18T10:38:59Z | https://web.archive.org/web/20260318103859/https://developers.openai.com/api/docs/pricing |
| gpt-5.4-nano-2026-03-17 | 2026-03-17 | ID | $0.20 | $1.25 | gpt-5.4-nano | yes | $0.20 | $1.25 | wayback-pricing | 2026-03-18T10:38:59Z | https://web.archive.org/web/20260318103859/https://developers.openai.com/api/docs/pricing |
| gpt-5.5-2026-04-23 | 2026-04-23 | ID | $5.00 | $30.00 | gpt-5.5 (<272K context length) | yes | $5.00 | $30.00 | wayback-pricing | 2026-04-25T10:46:44Z | https://web.archive.org/web/20260425104644/https://developers.openai.com/api/docs/pricing |
| gpt-5.6-luna | 2026-06-23 | API created | $0.20 | $1.20 | gpt-5.6-luna | yes | $1.00 | $6.00 | wayback-pricing | 2026-07-09T17:50:29Z | https://web.archive.org/web/20260709175029/https://developers.openai.com/api/docs/pricing |
| gpt-5.6-sol | 2026-06-23 | API created | $4.00 | $20.00 | gpt-5.6-sol | yes | $5.00 | $30.00 | wayback-pricing | 2026-07-09T17:50:29Z | https://web.archive.org/web/20260709175029/https://developers.openai.com/api/docs/pricing |
| gpt-5.6-terra | 2026-06-23 | API created | $2.00 | $12.00 | gpt-5.6-terra | yes | $2.50 | $15.00 | wayback-pricing | 2026-07-09T17:50:29Z | https://web.archive.org/web/20260709175029/https://developers.openai.com/api/docs/pricing |
| gpt-6-astra | 2026-08-27 | API created | $10.00 | $50.00 | gpt-6-astra | yes | $10.00 | $50.00 | wayback-pricing | 2026-09-04T10:30:51Z | https://web.archive.org/web/20260904103051/https://developers.openai.com/api/docs/pricing |
| gpt-6-luna | 2026-09-14 | API created | $0.10 | $0.50 | gpt-6-luna | yes | $0.10 | $0.50 | changelog | 2026-09-22 | https://developers.openai.com/api/docs/changelog.md |
| gpt-6-sol | 2026-09-14 | API created | $2.00 | $10.00 | gpt-6-sol | yes | $2.00 | $10.00 | changelog | 2026-09-22 | https://developers.openai.com/api/docs/changelog.md |

What moved between launch and today, from the same receipts: `o3` launched at $10 / $40 and
was cut to $2 / $8 on 2025-06-10; `gpt-5.6-terra` ($2.50 / $15 at launch) and `gpt-5.6-luna`
($1.00 / $6.00 at launch) were cut on 2026-07-30; `gpt-5.6-sol` lists at $5 / $30 and is on a
promotion at $4 / $20 "at least through November 21, 2026". Every dollar figure in the result
tables is at **today's** price, because that is what was billed.

## The projection, written before the full runs

Each candidate answered the first five six-way items (`pilot/`). The projection scales each
model's own measured tokens per prompt character to the whole task, and its own mean output
tokens to the item count, times a safety factor (1.0 without reasoning; 2.0 on task c and
3.0 on task d with reasoning). The roster was then fixed in `receipts/roster.json` before any
full run: phase A (the `gpt-4-0613` anchor, then one flagship per generation), B (the R-11
pair at default effort), C (the other generations and the small rungs), D and E (the
doorman, flagships first). `o1-2024-12-17` was held back by a rule written in advance: run
its six-way only if actual spend after E plus 1.5 × its pilot-mean projection ($3.46) stays
at or under $11.50. Phases D and E were then HELD (above), and the lead ordered the freed
budget kept for an operator's ⧖-5 ruling (the anchor sweep top-up) rather than spent on other
models, so phase F was not run either.

| model | reasoning | effort | pilot items | pilot correct | in tok / item | out tok / item | reasoning tok / item | reasoning tok max | pilot $ | projected c (108) $ | projected d (48) $ | run c | run d |
|---|---|---|---:|---:|---:|---:|---:|---:|---:|---:|---:|---|---|
| gpt-3.5-turbo-0125 | no | — | 5 | 1 | 293 | 7.8 | 0.0 | 0 | $0.0008 | $0.0162 | $0.0122 | yes | HELD (⧖-4) |
| gpt-3.5-turbo-1106 | no | — | 5 | 1 | 293 | 7.8 | 0.0 | 0 | $0.0015 | $0.0316 | $0.0239 | yes | HELD (⧖-4) |
| gpt-4-0613 | no | — | 5 | 4 | 293 | 2.8 | 0.0 | 0 | $0.0448 | $0.9154 | $0.7040 | yes | HELD (⧖-4) |
| gpt-4-turbo-2024-04-09 | no | — | 5 | 3 | 293 | 6.2 | 0.0 | 0 | $0.0156 | $0.3192 | $0.2409 | yes | HELD (⧖-4) |
| gpt-4.1-2025-04-14 | no | — | 5 | 5 | 284 | 3.2 | 0.0 | 0 | $0.0030 | $0.0608 | $0.0463 | yes | HELD (⧖-4) |
| gpt-4.1-mini-2025-04-14 | no | — | 5 | 3 | 284 | 4.8 | 0.0 | 0 | $0.0006 | $0.0124 | $0.0094 | yes | HELD (⧖-4) |
| gpt-4.1-nano-2025-04-14 | no | — | 5 | 1 | 284 | 2.0 | 0.0 | 0 | $0.0001 | $0.0030 | $0.0023 | yes | HELD (⧖-4) |
| gpt-4o-2024-05-13 | no | — | 5 | 3 | 284 | 4.0 | 0.0 | 0 | $0.0074 | $0.1516 | $0.1155 | yes | HELD (⧖-4) |
| gpt-4o-2024-08-06 | no | — | 5 | 5 | 284 | 5.0 | 0.0 | 0 | $0.0038 | $0.0780 | $0.0587 | yes | HELD (⧖-4) |
| gpt-4o-2024-11-20 | no | — | 5 | 3 | 284 | 2.4 | 0.0 | 0 | $0.0037 | $0.0752 | $0.0574 | yes | HELD (⧖-4) |
| gpt-4o-mini-2024-07-18 | no | — | 5 | 2 | 284 | 7.6 | 0.0 | 0 | $0.0002 | $0.0048 | $0.0036 | yes | HELD (⧖-4) |
| gpt-5-2025-08-07 | yes | minimal | 5 | 5 | 283 | 10.4 | 0.0 | 0 | $0.0023 | $0.0586 | $0.0430 | yes | HELD (⧖-4) |
| gpt-5-mini-2025-08-07 | yes | minimal | 5 | 4 | 283 | 10.6 | 0.0 | 0 | $0.0005 | $0.0118 | $0.0087 | yes | HELD (⧖-4) |
| gpt-5-nano-2025-08-07 | yes | minimal | 5 | 1 | 283 | 11.8 | 0.0 | 0 | $0.0001 | $0.0025 | $0.0018 | yes | HELD (⧖-4) |
| gpt-5.1-2025-11-13 | yes | none | 5 | 5 | 283 | 10.6 | 0.0 | 0 | $0.0023 | $0.0591 | $0.0433 | yes | HELD (⧖-4) |
| gpt-5.2-2025-12-11 | yes | none | 5 | 5 | 283 | 4.8 | 0.0 | 0 | $0.0028 | $0.0651 | $0.0489 | yes | HELD (⧖-4) |
| gpt-5.4-2026-03-05 | yes | none | 5 | 5 | 283 | 4.8 | 0.0 | 0 | $0.0039 | $0.0879 | $0.0665 | yes | HELD (⧖-4) |
| gpt-5.4-mini-2026-03-17 | yes | none | 5 | 3 | 283 | 10.4 | 0.0 | 0 | $0.0013 | $0.0318 | $0.0236 | yes | HELD (⧖-4) |
| gpt-5.4-nano-2026-03-17 | yes | none | 5 | 2 | 283 | 4.4 | 0.0 | 0 | $0.0003 | $0.0070 | $0.0053 | yes | HELD (⧖-4) |
| gpt-5.5-2026-04-23 | yes | none | 5 | 5 | 283 | 4.0 | 0.0 | 0 | $0.0077 | $0.1706 | $0.1295 | yes | HELD (⧖-4) |
| gpt-5.5-2026-04-23@default | yes | default (server; no reasoning_effort sent) | 5 | 5 | 283 | 69.0 | 59.0 | 81 | $0.0174 | $0.5918 | $0.4103 | yes | no |
| gpt-5.6-luna | yes | none | 5 | 5 | 283 | 4.4 | 0.0 | 0 | $0.0003 | $0.0069 | $0.0052 | yes | HELD (⧖-4) |
| gpt-5.6-sol | yes | none | 5 | 5 | 283 | 4.0 | 0.0 | 0 | $0.0061 | $0.1330 | $0.1013 | yes | HELD (⧖-4) |
| gpt-5.6-terra | yes | none | 5 | 5 | 283 | 4.0 | 0.0 | 0 | $0.0031 | $0.0682 | $0.0518 | yes | HELD (⧖-4) |
| gpt-6-astra | yes | low | 5 | 5 | 283 | 4.0 | 0.0 | 0 | $0.0152 | $0.3325 | $0.2532 | yes | HELD (⧖-4) |
| gpt-6-astra@default | yes | default (server; no reasoning_effort sent) | 5 | 5 | 283 | 4.0 | 0.0 | 0 | $0.0152 | $0.3325 | $0.2532 | yes | no |
| gpt-6-luna | yes | none | 5 | 5 | 283 | 4.4 | 0.0 | 0 | $0.0002 | $0.0034 | $0.0026 | yes | HELD (⧖-4) |
| gpt-6-sol | yes | none | 5 | 5 | 283 | 4.0 | 0.0 | 0 | $0.0030 | $0.0665 | $0.0506 | yes | HELD (⧖-4) |
| o1-2024-12-17 | yes | low | 5 | 5 | 283 | 463.0 | 448.0 | 768 | $0.1602 | $6.4344 | $4.3369 | no | no |
| o3-2025-04-16 | yes | low | 5 | 5 | 283 | 95.8 | 76.8 | 192 | $0.0067 | $0.2234 | $0.1552 | yes | HELD (⧖-4) |
| o3-mini-2025-01-31 | yes | low | 5 | 2 | 283 | 347.0 | 332.8 | 576 | $0.0092 | $0.3616 | $0.2445 | yes | HELD (⧖-4) |
| o4-mini-2025-04-16 | yes | low | 5 | 4 | 283 | 288.4 | 268.8 | 512 | $0.0079 | $0.3059 | $0.2074 | yes | HELD (⧖-4) |

**`o1-2024-12-17`: not run beyond its five-item pilot (5 of 5, $0.1602).** Its
six-way projected $3.46 at the pilot's mean and $6.43 with the safety factor, and the
phase-F rule was overtaken by the hold. It is first in line for the anchor sweep (⧖-5), whose
priced plan is `receipts/sweep-priced.json`.

## Results: the six-way (task c), in release order

| model | date | resolved snapshot | effort | temperature | scored | correct | accuracy | 95 % low | 95 % high | strict correct | format failures | failed calls | truncated | dollars (today) | dollars / correct (today) | dollars at launch price | dollars / correct at launch price |
|---|---|---|---|---|---:|---:|---:|---:|---:|---:|---:|---:|---:|---:|---:|---:|---:|
| gpt-4-0613 | 2023-06-13 | gpt-4-0613 | n/a (no reasoning) | 0 | 108 | 88 | 81.5 % | 73.1 % | 87.7 % | 88 | 0 | 0 | 0 | $0.9553 | $0.01086 | $0.9553 | $0.01086 |
| gpt-3.5-turbo-1106 | 2023-11-06 | gpt-3.5-turbo-1106 | n/a (no reasoning) | 0 | 108 | 34 | 31.5 % | 23.5 % | 40.7 % | 34 | 0 | 0 | 0 | $0.0322 | $0.00095 | $0.0322 | $0.00095 |
| gpt-3.5-turbo-0125 | 2024-01-25 | gpt-3.5-turbo-0125 | n/a (no reasoning) | 0 | 108 | 24 | 22.2 % | 15.4 % | 30.9 % | 24 | 0 | 0 | 0 | $0.0165 | $0.00069 | $0.0165 | $0.00069 |
| gpt-4-turbo-2024-04-09 | 2024-04-09 | gpt-4-turbo-2024-04-09 | n/a (no reasoning) | 0 | 108 | 92 | 85.2 % | 77.3 % | 90.7 % | 92 | 0 | 0 | 0 | $0.3280 | $0.00357 | $0.3280 | $0.00357 |
| gpt-4o-2024-05-13 | 2024-05-13 | gpt-4o-2024-05-13 | n/a (no reasoning) | 0 | 108 | 88 | 81.5 % | 73.1 % | 87.7 % | 88 | 0 | 0 | 0 | $0.1551 | $0.00176 | $0.1551 | $0.00176 |
| gpt-4o-mini-2024-07-18 | 2024-07-18 | gpt-4o-mini-2024-07-18 | n/a (no reasoning) | 0 | 108 | 73 | 67.6 % | 58.3 % | 75.7 % | 73 | 0 | 0 | 0 | $0.0049 | $0.00007 | $0.0049 | $0.00007 |
| gpt-4o-2024-08-06 | 2024-08-06 | gpt-4o-2024-08-06 | n/a (no reasoning) | 0 | 108 | 90 | 83.3 % | 75.2 % | 89.2 % | 90 | 0 | 0 | 0 | $0.0809 | $0.00090 | $0.0809 | $0.00090 |
| gpt-4o-2024-11-20 | 2024-11-20 | gpt-4o-2024-11-20 | n/a (no reasoning) | 0 | 108 | 89 | 82.4 % | 74.2 % | 88.4 % | 89 | 0 | 0 | 0 | $0.0768 | $0.00086 | $0.0768 | $0.00086 |
| o3-mini-2025-01-31 | 2025-01-31 | o3-mini-2025-01-31 | low | default (1) | 108 | 79 | 73.1 % | 64.1 % | 80.6 % | 79 | 0 | 0 | 0 | $0.1445 | $0.00183 | $0.1445 | $0.00183 |
| gpt-4.1-2025-04-14 | 2025-04-14 | gpt-4.1-2025-04-14 | n/a (no reasoning) | 0 | 108 | 97 | 89.8 % | 82.7 % | 94.2 % | 97 | 0 | 0 | 0 | $0.0631 | $0.00065 | $0.0631 | $0.00065 |
| gpt-4.1-mini-2025-04-14 | 2025-04-14 | gpt-4.1-mini-2025-04-14 | n/a (no reasoning) | 0 | 108 | 84 | 77.8 % | 69.1 % | 84.6 % | 84 | 0 | 0 | 0 | $0.0127 | $0.00015 | $0.0127 | $0.00015 |
| gpt-4.1-nano-2025-04-14 | 2025-04-14 | gpt-4.1-nano-2025-04-14 | n/a (no reasoning) | 0 | 108 | 41 | 38.0 % | 29.4 % | 47.4 % | 41 | 0 | 0 | 0 | $0.0031 | $0.00008 | $0.0031 | $0.00008 |
| o3-2025-04-16 | 2025-04-16 | o3-2025-04-16 | low | default (1) | 108 | 98 | 90.7 % | 83.8 % | 94.9 % | 98 | 0 | 0 | 0 | $0.1975 | $0.00202 | $0.9876 | $0.01008 |
| o4-mini-2025-04-16 | 2025-04-16 | o4-mini-2025-04-16 | low | default (1) | 108 | 93 | 86.1 % | 78.3 % | 91.4 % | 93 | 0 | 0 | 0 | $0.1433 | $0.00154 | $0.1433 | $0.00154 |
| gpt-5-2025-08-07 | 2025-08-07 | gpt-5-2025-08-07 | minimal | default (1) | 108 | 97 | 89.8 % | 82.7 % | 94.2 % | 97 | 0 | 0 | 0 | $0.0484 | $0.00050 | $0.0484 | $0.00050 |
| gpt-5-mini-2025-08-07 | 2025-08-07 | gpt-5-mini-2025-08-07 | minimal | default (1) | 108 | 86 | 79.6 % | 71.1 % | 86.1 % | 86 | 0 | 0 | 0 | $0.0097 | $0.00011 | $0.0097 | $0.00011 |
| gpt-5-nano-2025-08-07 | 2025-08-07 | gpt-5-nano-2025-08-07 | minimal | default (1) | 108 | 45 | 41.7 % | 32.8 % | 51.1 % | 45 | 0 | 0 | 0 | $0.0019 | $0.00004 | $0.0019 | $0.00004 |
| gpt-5.1-2025-11-13 | 2025-11-13 | gpt-5.1-2025-11-13 | none | 0 | 108 | 96 | 88.9 % | 81.6 % | 93.5 % | 96 | 0 | 0 | 0 | $0.0485 | $0.00050 | $0.0485 | $0.00050 |
| gpt-5.2-2025-12-11 | 2025-12-11 | gpt-5.2-2025-12-11 | none | 0 | 108 | 96 | 88.9 % | 81.6 % | 93.5 % | 96 | 0 | 0 | 0 | $0.0592 | $0.00062 | $0.0592 | $0.00062 |
| gpt-5.4-2026-03-05 | 2026-03-05 | gpt-5.4-2026-03-05 | none | 0 | 108 | 98 | 90.7 % | 83.8 % | 94.9 % | 98 | 0 | 0 | 0 | $0.0815 | $0.00083 | $0.0815 | $0.00083 |
| gpt-5.4-mini-2026-03-17 | 2026-03-17 | gpt-5.4-mini-2026-03-17 | none | 0 | 108 | 90 | 83.3 % | 75.2 % | 89.2 % | 90 | 0 | 0 | 0 | $0.0270 | $0.00030 | $0.0270 | $0.00030 |
| gpt-5.4-nano-2026-03-17 | 2026-03-17 | gpt-5.4-nano-2026-03-17 | none | 0 | 108 | 52 | 48.1 % | 39.0 % | 57.5 % | 52 | 0 | 0 | 0 | $0.0065 | $0.00013 | $0.0065 | $0.00013 |
| gpt-5.5-2026-04-23 | 2026-04-23 | gpt-5.5-2026-04-23 | none | 0 | 108 | 102 | 94.4 % | 88.4 % | 97.4 % | 102 | 0 | 0 | 0 | $0.1609 | $0.00158 | $0.1609 | $0.00158 |
| gpt-5.5-2026-04-23@default | 2026-04-23 | gpt-5.5-2026-04-23 | default = medium (server echo) | default (1) | 108 | 103 | 95.4 % | 89.6 % | 98.0 % | 103 | 0 | 0 | 0 | $0.4691 | $0.00455 | $0.4691 | $0.00455 |
| gpt-5.6-luna | 2026-06-23 | gpt-5.6-luna | none | 0 | 108 | 94 | 87.0 % | 79.4 % | 92.1 % | 94 | 0 | 0 | 0 | $0.0065 | $0.00007 | $0.0326 | $0.00035 |
| gpt-5.6-sol | 2026-06-23 | gpt-5.6-sol | none | 0 | 108 | 103 | 95.4 % | 89.6 % | 98.0 % | 103 | 0 | 0 | 0 | $0.1270 | $0.00123 | $0.1609 | $0.00156 |
| gpt-5.6-terra | 2026-06-23 | gpt-5.6-terra | none | 0 | 108 | 95 | 88.0 % | 80.5 % | 92.8 % | 95 | 0 | 0 | 0 | $0.0647 | $0.00068 | $0.0809 | $0.00085 |
| gpt-6-astra | 2026-08-27 | gpt-6-astra | low | default (1) | 108 | 101 | 93.5 % | 87.2 % | 96.8 % | 101 | 0 | 0 | 0 | $0.3327 | $0.00329 | $0.3327 | $0.00329 |
| gpt-6-astra@default | 2026-08-27 | gpt-6-astra | default = medium (server echo) | default (1) | 108 | 101 | 93.5 % | 87.2 % | 96.8 % | 101 | 0 | 0 | 0 | $0.3474 | $0.00344 | $0.3474 | $0.00344 |
| gpt-6-luna | 2026-09-14 | gpt-6-luna | none | 0 | 108 | 96 | 88.9 % | 81.6 % | 93.5 % | 96 | 0 | 0 | 0 | $0.0032 | $0.00003 | $0.0032 | $0.00003 |
| gpt-6-sol | 2026-09-14 | gpt-6-sol | none | 0 | 108 | 101 | 93.5 % | 87.2 % | 96.8 % | 101 | 0 | 0 | 0 | $0.0637 | $0.00063 | $0.0637 | $0.00063 |

### Timings for the six-way

All network-included, from the laptop to api.openai.com. TTFT is the arrival of the first
streamed event carrying a visible token; for a reasoning model that includes its hidden
reasoning. "stream out tok/s" divides the completion tokens by the span between the first and
last streamed events: a streaming rate, not a server decode speed, and for a one-letter reply
it is noise. OpenAI returns no server-side prefill or decode timing, so none is claimed.

| model | timing | TTFT median | TTFT p95 | total median | total p95 | tokens in | tokens out | of which reasoning | visible out | reasoning max / item | stream out tok/s median |
|---|---|---:|---:|---:|---:|---:|---:|---:|---:|---:|---:|
| gpt-4-0613 | streamed, network-included | 1.38 s | 2.62 s | 1.51 s | 2.70 s | 30544 | 650 | 0 | 650 | 0 | 41.9 |
| gpt-3.5-turbo-1106 | streamed, network-included | 1.66 s | 3.61 s | 1.77 s | 3.70 s | 30544 | 837 | 0 | 837 | 0 | 85.3 |
| gpt-3.5-turbo-0125 | streamed, network-included | 1.46 s | 7.42 s | 1.59 s | 7.54 s | 30544 | 844 | 0 | 844 | 0 | 75.5 |
| gpt-4-turbo-2024-04-09 | streamed, network-included | 2.07 s | 9.34 s | 2.21 s | 9.46 s | 30544 | 753 | 0 | 753 | 0 | 33.4 |
| gpt-4o-2024-05-13 | streamed, network-included | 1.20 s | 6.66 s | 1.27 s | 6.73 s | 29701 | 437 | 0 | 437 | 0 | 52.0 |
| gpt-4o-mini-2024-07-18 | streamed, network-included | 0.92 s | 4.90 s | 1.02 s | 4.95 s | 29701 | 754 | 0 | 754 | 0 | 69.5 |
| gpt-4o-2024-08-06 | streamed, network-included | 1.13 s | 6.73 s | 1.21 s | 6.84 s | 29701 | 663 | 0 | 663 | 0 | 63.4 |
| gpt-4o-2024-11-20 | streamed, network-included | 1.03 s | 6.65 s | 1.08 s | 6.69 s | 29701 | 251 | 0 | 251 | 0 | 36.2 |
| o3-mini-2025-01-31 | streamed, network-included | 2.05 s | 4.72 s | 2.05 s | 4.72 s | 29593 | 25433 | 23872 | 1561 | 1024 | 3604.1 |
| gpt-4.1-2025-04-14 | streamed, network-included | 0.84 s | 1.90 s | 0.88 s | 1.95 s | 29701 | 468 | 0 | 468 | 0 | 61.2 |
| gpt-4.1-mini-2025-04-14 | streamed, network-included | 1.12 s | 3.09 s | 1.20 s | 3.33 s | 29701 | 500 | 0 | 500 | 0 | 55.4 |
| gpt-4.1-nano-2025-04-14 | streamed, network-included | 1.21 s | 3.19 s | 1.24 s | 3.23 s | 29701 | 274 | 0 | 274 | 0 | 35.0 |
| o3-2025-04-16 | streamed, network-included | 2.65 s | 6.71 s | 2.68 s | 6.76 s | 29593 | 17292 | 15232 | 2060 | 1088 | 1549.8 |
| o4-mini-2025-04-16 | streamed, network-included | 3.31 s | 7.25 s | 3.32 s | 7.32 s | 29593 | 25163 | 23104 | 2059 | 704 | 3158.9 |
| gpt-5-2025-08-07 | streamed, network-included | 1.74 s | 5.78 s | 1.76 s | 5.85 s | 29593 | 1144 | 0 | 1144 | 0 | 138.7 |
| gpt-5-mini-2025-08-07 | streamed, network-included | 1.50 s | 2.38 s | 1.55 s | 2.40 s | 29593 | 1163 | 0 | 1163 | 0 | 139.5 |
| gpt-5-nano-2025-08-07 | streamed, network-included | 1.47 s | 3.91 s | 1.51 s | 3.96 s | 29593 | 1138 | 0 | 1138 | 0 | 175.3 |
| gpt-5.1-2025-11-13 | streamed, network-included | 0.96 s | 2.47 s | 0.99 s | 2.54 s | 29593 | 1148 | 0 | 1148 | 0 | 152.8 |
| gpt-5.2-2025-12-11 | streamed, network-included | 1.19 s | 3.10 s | 1.25 s | 3.18 s | 29593 | 526 | 0 | 526 | 0 | 63.5 |
| gpt-5.4-2026-03-05 | streamed, network-included | 1.11 s | 5.24 s | 1.16 s | 5.31 s | 29593 | 501 | 0 | 501 | 0 | 77.5 |
| gpt-5.4-mini-2026-03-17 | streamed, network-included | 0.91 s | 6.23 s | 0.95 s | 6.28 s | 29593 | 1068 | 0 | 1068 | 0 | 173.8 |
| gpt-5.4-nano-2026-03-17 | streamed, network-included | 0.90 s | 4.17 s | 0.93 s | 4.19 s | 29593 | 468 | 0 | 468 | 0 | 75.8 |
| gpt-5.5-2026-04-23 | streamed, network-included | 1.35 s | 4.92 s | 1.37 s | 4.93 s | 29593 | 432 | 0 | 432 | 0 | 71.9 |
| gpt-5.5-2026-04-23@default | streamed, network-included | 2.78 s | 8.33 s | 2.83 s | 8.36 s | 29593 | 10706 | 9625 | 1081 | 512 | 1298.1 |
| gpt-5.6-luna | streamed, network-included | 1.28 s | 3.39 s | 1.29 s | 3.42 s | 29593 | 494 | 0 | 494 | 0 | 92.5 |
| gpt-5.6-sol | streamed, network-included | 1.41 s | 3.89 s | 1.41 s | 3.89 s | 29593 | 432 | 0 | 432 | 0 | 101.7 |
| gpt-5.6-terra | streamed, network-included | 1.39 s | 6.19 s | 1.39 s | 6.21 s | 29593 | 458 | 0 | 458 | 0 | 91.2 |
| gpt-6-astra | streamed, network-included | 1.94 s | 9.12 s | 1.94 s | 9.12 s | 29593 | 736 | 232 | 504 | 39 | 102.7 |
| gpt-6-astra@default | streamed, network-included | 2.00 s | 8.34 s | 2.00 s | 8.34 s | 29593 | 1030 | 466 | 564 | 78 | 104.0 |
| gpt-6-luna | streamed, network-included | 0.92 s | 3.72 s | 0.95 s | 3.75 s | 29593 | 479 | 0 | 479 | 0 | 89.8 |
| gpt-6-sol | streamed, network-included | 1.15 s | 2.42 s | 1.17 s | 2.47 s | 29593 | 452 | 0 | 452 | 0 | 97.1 |

## Reasoning on and off (R-11)

The same model at the lowest effort it accepts and at the server's own default. The default's
name was read from the server, not assumed: a Responses API call with no effort sent echoes
the effort it used (`receipts/default-effort/`). At its default effort `gpt-5.5` refuses
temperature 0 (the refused pilot is kept at
`raw/pilot/gpt-5.5-2026-04-23@default.temperature0-refused/`), so its default row also runs at
the default temperature: the two rows differ in effort and temperature, not effort alone.

| model | effort | temperature | scored | correct | accuracy | 95 % low | 95 % high | reasoning tokens (all items) | reasoning max / item | TTFT median | dollars | dollars / correct |
|---|---|---|---:|---:|---:|---:|---:|---:|---:|---:|---:|---:|
| gpt-5.5-2026-04-23 | none | 0 | 108 | 102 | 94.4 % | 88.4 % | 97.4 % | 0 | 0 | 1.35 s | $0.1609 | $0.00158 |
| gpt-5.5-2026-04-23 | default = medium (server echo) | default (1) | 108 | 103 | 95.4 % | 89.6 % | 98.0 % | 9625 | 512 | 2.78 s | $0.4691 | $0.00455 |
| gpt-6-astra | low | default (1) | 108 | 101 | 93.5 % | 87.2 % | 96.8 % | 232 | 39 | 1.94 s | $0.3327 | $0.00329 |
| gpt-6-astra | default = medium (server echo) | default (1) | 108 | 101 | 93.5 % | 87.2 % | 96.8 % | 466 | 78 | 2.00 s | $0.3474 | $0.00344 |

**What reasoning bought on this task: about one item.** `gpt-5.5` at its default
(medium) effort scored 103 against 102 at `none`, spent 9,625 hidden reasoning tokens (at
most 512 on any item), and cost 2.9 times as much ($0.4691 against $0.1609). Its median
TTFT roughly doubled, from 1.35 s to 2.78 s. `gpt-6-astra` accepts no effort lower than
`low`, and it barely reasons on this task at either setting: 232 reasoning tokens in total
at `low` and 466 at medium, both 101 of 108. On a one-letter decision task these models do
not need to think. R-8's quality family is where the price of thinking should show.

## The doorman planted set (task d): HELD on ⧖-4

Not read. The adapter, the projection (above, "projected d") and the phase order are in
place; if ⧖-4 rules the set may leave the estate, creating `../DOORMAN-EGRESS-RULED` lets
`run_phases.sh`'s phases D and E run as written, flagships first, and `tables.py` has the
results and timing tables for it (`results_d`, `timing_d`) ready. The local rows it would be
read beside are in `jev-2026-09-21/rows-addenda/` (gemma4 FP8 generate 35 of 48; OpenJev FP8
generate 45 of 48).

## Beside the local rows

The local comparators, read from their own report files in the Jev bench (never retyped).
"strict correct" recounts each file's rows with an unparsed reply as wrong. The Wilson
intervals are computed here from the same counts.

| arm | where | scored | correct | accuracy | 95 % low | 95 % high | strict correct | unparsed | median latency | run started (UTC) | source file |
|---|---|---:|---:|---:|---:|---:|---:|---:|---:|---|---|
| gemma4:26b generate (ollama Q4_K_M), the Jev base-generate arm | benchbox, one RTX 3090 | 108 | 89 | 82.4 % | 74.2 % | 88.4 % | 88 | 1 | 1.06 s | 2026-09-21T17:56:12Z | `jev-2026-09-21/rows/base-generate.c.report.json` |
| gemma4 26B-A4B FP8 generate (vLLM), addenda arm 10 | benchbox, two RTX 3090s (TP 2) | 108 | 89 | 82.4 % | 74.2 % | 88.4 % | 88 | 2 | 0.06 s | 2026-09-21T22:01:31Z | `jev-2026-09-21/rows-addenda/gemma4-fp8-generate.c.report.json` |
| OpenJev FP8 generate (vLLM), arm 2 | benchbox, two RTX 3090s (TP 2) | 108 | 101 | 93.5 % | 87.2 % | 96.8 % | 101 | 0 | 0.39 s | 2026-09-21T19:00:34Z | `jev-2026-09-21/rows/openjev-fp8-generate.c.report.json` |
| OpenJev FP8 readout (vLLM), arm 2 | benchbox, two RTX 3090s (TP 2) | 108 | 101 | 93.5 % | 87.2 % | 96.8 % | 101 | 0 | 0.35 s | 2026-09-21T18:48:14Z | `jev-2026-09-21/rows/openjev-fp8-readout.c.report.json` |

**Read beside the OpenAI rows:** benchbox's `gemma4:26b` (one 3090, ollama
Q4_K_M) sits on `gpt-4-0613`'s score (89 against 88, or 88 against 88 strict) and on
`gpt-4o-2024-05-13`'s (88). OpenJev FP8 (two 3090s, vLLM, CC BY-NC, bench-only) sits on
`gpt-6-astra` and `gpt-6-sol` (101). No local row reaches `gpt-5.6-sol`'s 103. These are the
same 108 items through the same prompt builder and parse. The local rows are dated
2026-09-21 and the OpenAI rows 2026-09-23. The two use different transports (ollama/vLLM on
the LAN, the public API over the internet), so the latency columns are not comparable
across them. The accuracy columns are.

## Caveats

- **Effort is not uniform, because the models do not allow it to be.** `gpt-5.1` and
  later (except `gpt-6-astra`) ran at `none`, i.e. with no reasoning. `gpt-5`, `-mini` and
  `-nano` ran at `minimal`. The o-series and `gpt-6-astra` ran at `low`, because `none` and
  `minimal` are refused (`receipts/profiles/`). Every row carries its effort.
- **Temperature 0 where accepted, the default (1) elsewhere.** The o-series, `gpt-5`,
  `-mini`, `-nano`, `gpt-6-astra` and both `@default` rows refuse 0 and ran sampled. Each
  (model, task) was read once, so a sampled row's score can move by an item or two on a
  re-read. `gpt-4-0613` returns no `system_fingerprint` at all. PLAN-v2's
  determinism repeat (the anchor sweep) is the receipt for that.
- **The token limit differs by class** (above): 8 visible tokens without reasoning, as the
  local arm had; a 4,096 budget with reasoning. No row was truncated and no call failed, so
  the artefact rules changed no figure, and every `strict` count equals its harness count
  (no unparsed replies).
- **n = 108, one reading each.** The Wilson intervals are about ±7 points at 90 %.
  Differences of one to three items (the whole of the gpt-4.1 → gpt-6 stretch) are not
  separable on this set. PLAN-v2's paired tests (R-15) are the instrument for "matches".
- **Contamination (R-1): clean by construction on the evidence available.** Every id read
  here was created on the API by 2026-09-14, and task_c's lines first went public on
  2026-09-19 (the long table's wall). The undated ids (`gpt-5.6-*`, `gpt-6-*`) could be
  re-pointed later behind the same name. The response's `model` field matched the asked id
  on every row, and the fingerprints are kept.
- **Home-team text (PLAN-v2 §2.1):** task_c's lines were written by gemma-4-26B or
  mistral-small-3.2, which favours the local gemma arm if it favours anything.
- **Dates.** "date" is the id's own snapshot date, or the API `created` date for undated ids.
  OpenAI's public release dates for the 2026 undated ids are later (prices table note).
- **Latency is network-included from the laptop** (a laptop on a home line), with p95 tails to
  about 9 s. The TTFT columns are this path's, not OpenAI's server times.
- **`gpt-5.3-chat-latest`** is listed by `/v1/models` but refused as deprecated. It is
  unreadable, so there is no 5.3 point. **`o1`** was piloted only (above).
- **The doorman planted set is HELD on ⧖-4**, and no doorman item left the estate from
  this lane. When it is ruled, only its 36 in-contract lines may go
  (`test_doorman_egress.py`: 7 tests, and they fail if either fence is removed).
- **Protocol P0** in PLAN-v2's terms. P1's reclassification of `truncated` as
  `budget_exhausted` (counted wrong) changes nothing here, because there were none.

## Spend

Every successful call's dollars, from its own usage block at today's price
(`receipts/spend.jsonl`). The cap was $12.00 of a $17.14 balance.

| stage | task | calls | dollars |
|---|---|---:|---:|
| pilot | c | 160 | $0.3470 |
| probe | default-effort | 2 | $0.0013 |
| probe | probe | 30 | $0.0074 |
| rows | c | 3348 | $4.0719 |
| total |  | 3540 | $4.4277 |

**Total spent by this lane: $4.4277.**

## Files

- `run_openai.py`: the adapter (probe, pilot, rows); `oai.py`: transport, prices, ledger;
  `default_effort.py`: R-11's server echo; `run_phases.sh`: phases A–E in order.
- `project.py` → `receipts/projection.json`; `receipts/roster.json`: the roster decision.
- `tables.py` + `report.template.txt` → this file (`python3 -B tables.py --write`).
- `rows/<model>.<task>.jsonl` and `.report.json`: every row and its summary; `pilot/`: the
  five-item pilot; `raw/<stage>/<model>/<task>/`: every raw response with its streamed events,
  named by prompt sha.
- `receipts/`: the model list, the pricing page, the probes (`profiles/`), the default-effort
  echoes, the launch-price pages, the logs, and `spend.jsonl`.