Skip to content
Thomas Fairhurst
Chapters

Chapter 11 · Models

Fable decides, Opus builds, GPT reads

I gave the strongest model the decisions, a cheaper one the work a test can check, and a second vendor's model the lost games and the code reviews. 7,720 agent threads ran on flat-rate subscriptions; one 72-hour stretch alone would have cost about $7,900 at list price.

7,720agent threads, 26 Oct to 13 Nov
73agents running at once at the peak, 6 Nov
3.8weeks of one subscription's usage, drawn in 72 hours

Each model on the work it does best

Opus 5.5 did most of the work: building, fixing and tracing, work that tests and games can check. Fable, the strongest and dearest model, directed. GPT, from another vendor, read lost games and reviewed code.

Agent threads by model

Opus 5.5builds, fixes, traces lost games, reviews
4,617
GPT‑6.1 Solreads lost games, reviews code
1,879
Haiku 5.5relays: each starts one GPT review
1,065
Fable 5.1 and 5.2orchestration: the Director, its leads and workflow steps
149
Otherstwo GPT‑6 models and Sonnet
10

Threads by period

PeriodOpus 5.5FableHaiku 5.5GPTOther
26 to 31 Oct, the first version564861––
1 to 4 Nov, planning and proofs2,692–1,0614555
5 to 9 Nov, Opus coordinates7161–1,4243
9 to 13 Nov, Fable directs645623–2
Each thread counts under the model it last ran on. Threads from 22 to 26 Oct, on the earlier hosts, were not kept.

A second vendor reads and reviews

GPT ran on a separate subscription through Codex, OpenAI’s command-line agent, so it added capacity without touching Claude’s limits. Agents started every GPT thread, between 3 and 7 Nov.

What the GPT threads did

Read a lost game204 on 5 Nov, 1,160 on 6 Nov
1,364
Reviewed code3 to 6 Nov
475
Audits, map-judge batches, tests
about 32
Night analyses
8
“I don't trust ChatGPT as much as I trust Opus.”
3 Nov
“Don't use Codex sub-agents to build stuff.”
5 Nov

GPT read and judged; Opus kept the final say and wrote the code. The 1,065 Haiku threads were relays, almost all on 3 Nov, because a review script could start only Claude agents.

  • Review script
  • Starts a Haiku agent
  • It runs one GPT review
  • And reads the result back
The relays retyped curly quotes, which broke a checksum and lost reviews. That evening builders started running their own reviews, and the relays went.

What a subscription-week buys

Claude ran on about seven flat-rate subscriptions, each with a weekly limit. Fitted on nine weeks that ran out, one login’s week bought:

List-price value of one login-week

Opus 5.5
about $3,200
Fable 5.2Fable is capped at half the week
about $550

Per list dollar, Fable drew 2.1 to 2.9 times what Opus did. So Fable got only the few directing threads, and from 10 Nov the gateway kept Opus off any login whose remaining week Fable needed.

Spending the limits on purpose

I treated usage as a budget: sized the agent count to what was left, moved work to whichever subscription had room, and refilled logins when the work was worth it.

  1. Ten sub-agents at a timeJudged “about the right amount to stay on pace” with the subscriptions.
  2. Fable 5.1 orchestrates, sub-agents stay on OpusThe strongest model where decisions are made, the cheaper one where tests check the work.
  3. Reviews move to GPT“We're running low on Opus and Claude inference, but we have a bunch of GPT inference.”
  4. GPT's subscription runs outReviews go back to Opus until it resets on 4 Nov, 08:52.
  5. GPT's subscription empty for goodThe last GPT thread had started a minute earlier.
  6. About $20 an agent-hourThe coordinator loop, priced at list. I ask to size the agent count to what is left.
  7. About 3.2 login-weeks leftAgainst a draw of 3.3 a day.
  8. Opus for every sub-agent“Fable is too expensive.” By that evening a standing rule.
  9. I supply the inferenceI held 7 banked resets, each refilling one login completely. Meters show refills on 11 and 12 Nov.
“Usage shouldn't be a worry. I'll use resets if needed.”
11 Nov

72 hours, priced

Only one stretch has a full token record: 9 Nov 23:33 to 12 Nov 23:25. At list price it comes to about $7,900, or 3.8 weeks of one subscription.

What the $7,900 paid for

Fable 5.2all token types
$2,900
Opus 5.5, reading cached context
$2,170
Opus 5.5, writing to the cache
$1,690
Opus 5.5, output
$1,130

Tokens in those 72 hours (M million, B billion)

Cached context read
10.85B
5.95B
Written to the cache
211M
56.5M
Output
56.7M
20.3M
  • Opus 5.5
  • Fable 5.2
The Fable thread that took over as Director on 10 Nov cost about $1,670 of it; the first Director thread, about $160. Before 9 Nov only fragments survive, so the whole project has no total.

Many agents at once

Unused inference is lost at each weekly reset, so before one I ran wide: up to 73 sub-agents at the same moment.

Most sub-agents running at the same moment

Any model6 Nov, 13:31: 67 GPT loss readers, 6 Opus
73
Claude only3 Nov, 13:40: 33 of them Haiku relays
59
Opus only3 Nov, 16:20
40

3 Nov, 16:56: “peaking to 42 agents is completely fine … we need to use that inference before the reset tomorrow.”

Long contexts, summarised on schedule

Opus and Fable hold up to a million tokens of context. Near a set threshold the harness summarises a thread and carries on, so a long-running agent never stops for a full context.

Settings and why

SettingValueWhy
Sub-agent cache lifetime1 hour, not 5 minutesfive-minute caches were 41% of sub-agent spend
Opus summary threshold250,000 tokenschosen from a replay of 790 summaries
Fable summary threshold500,000, then the full million from 12 Novthe same replay, then my call: “make it 1m”

The Director’s summaries

26summaries, each at about 468,000 tokens
27–47 minbetween two summaries
110–153 seach one took