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
Threads by period
| Period | Opus 5.5 | Fable | Haiku 5.5 | GPT | Other |
|---|---|---|---|---|---|
| 26 to 31 Oct, the first version | 564 | 86 | 1 | – | – |
| 1 to 4 Nov, planning and proofs | 2,692 | – | 1,061 | 455 | 5 |
| 5 to 9 Nov, Opus coordinates | 716 | 1 | – | 1,424 | 3 |
| 9 to 13 Nov, Fable directs | 645 | 62 | 3 | – | 2 |
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
“I don't trust ChatGPT as much as I trust Opus.”
“Don't use Codex sub-agents to build stuff.”
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
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
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.
- Ten sub-agents at a timeJudged “about the right amount to stay on pace” with the subscriptions.
- Fable 5.1 orchestrates, sub-agents stay on OpusThe strongest model where decisions are made, the cheaper one where tests check the work.
- Reviews move to GPT“We're running low on Opus and Claude inference, but we have a bunch of GPT inference.”
- GPT's subscription runs outReviews go back to Opus until it resets on 4 Nov, 08:52.
- GPT's subscription empty for goodThe last GPT thread had started a minute earlier.
- About $20 an agent-hourThe coordinator loop, priced at list. I ask to size the agent count to what is left.
- About 3.2 login-weeks leftAgainst a draw of 3.3 a day.
- Opus for every sub-agent“Fable is too expensive.” By that evening a standing rule.
- 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.”
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
Tokens in those 72 hours (M million, B billion)
- Opus 5.5
- Fable 5.2
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
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
| Setting | Value | Why |
|---|---|---|
| Sub-agent cache lifetime | 1 hour, not 5 minutes | five-minute caches were 41% of sub-agent spend |
| Opus summary threshold | 250,000 tokens | chosen from a replay of 790 summaries |
| Fable summary threshold | 500,000, then the full million from 12 Nov | the same replay, then my call: “make it 1m” |