Fable 5.1 or Opus 5, domain by domain
This settles which model a spawned specialist runs on, per domain, when Jarvis hands work to a sub-agent (a fresh, separate Claude with its own context). Every number is from Anthropic's two system cards, both run at their maximum effort setting; the Fable 5.1 numbers were produced with its safety filters switched on, exactly as it runs for you. Where a domain has no published test, the row says so rather than guessing. Read the two columns on the right first: the recommendation, and how solid it is.
- Click a column heading to sort. Click a row to expand its long cells. The buttons filter to the rows where something would change.
- Fable 5.1 means move that domain to Fable 5.1. Opus 5 means it stays where it is today.
- "How solid" is the honest strength of the evidence: solid (a direct test, same conditions), reasonable (a related test, consistent direction), thin (no test; reasoning by analogy).
- Two decisions are already yours from 1 September and are shown as such: legal stays on Opus 5; the cheap tier is tested first.
| Domain | What we spawn it for | The evidence (Fable 5.1 vs Opus 5) | The caveat | Recommendation | How solid, and what would flip it |
|---|
Cost, measured on your own usage rather than the price list
The price list says Fable 5.1 is double Opus 5 per fresh token ($10 and $50 per million against $5 and $25), and half on re-reading cached context ($0.25 against $0.50). Which of those dominates depends on the shape of the work, so both shapes were priced from your own transcripts of the last seven days.
- Your interactive sessions (207 sessions; 97.9% of tokens were cache re-reads): the same work costs 0.91x on Fable 5.1. Cheaper.
- Spawned sub-agents (295 agent runs; 94.8% cache re-reads, 5.0% fresh cache writes because each starts from a blank context): the same work costs 1.16x on Fable 5.1. Sixteen percent more, not double.
- These are list-price equivalents. You are on a subscription, so the figures describe how fast the allowance is used, not money billed.
The safety filter, in plain English, because it is the other half of the decision
Anthropic screens every request to Fable 5.1. When a request looks like it could be about hacking, the answer is quietly produced by Opus 4.8 instead (an older, weaker model). When it looks like it could be about biology, the answer is produced by Opus 5 (the same tier you use today, so no loss). Opus 5 has only the hacking screen, and it trips far less often. Legal, tax, finance and immigration questions do not look like hacking, so for those domains the filter is not the reason to avoid Fable 5.1; for security work it is the whole reason.
- On Anthropic's adversarial coding test, about half of Fable 5.1's coding runs were answered by Opus 4.8; under 10% of its tool-use and computer-use runs were. Its overall rate on that test was 23% (Fable 5 had been 60%). Card section 5.2.1, page 83.
- In your own sessions since 1 August: Opus 5 changed model mid-session once in 768 sessions; Fable 5 did so in 6 of 8; Fable 5.1 has run one session, too few to say.
- The figure that has been justifying "Opus 5 for irreversible domains" in the rules, "42% of calls on 26% of trials", was about Fable 5 on a benchmark the Fable 5.1 card does not report at all. It should not be quoted for Fable 5.1, and the rule text will be corrected whichever way you decide.
The three scheduled jobs that could move
- Telegram assistant (always on; answers your Telegram messages, updates contacts, runs the morning and evening briefings). On Opus 5. Nothing measured says it would improve on Fable 5.1, it is a general assistant workload with no domain test, and each message starts from a fresh context so the sub-agent cost shape (1.16x) applies. Recommendation: stay on Opus 5 for now.
- Eulia chat watch (every minute, five company chat rooms, decides whether anything needs acting on). On Opus 5. About 1,440 fresh calls a day, so cost shape matters most here, and there is no evidence of gain. Two other sessions are editing this file today. Recommendation: stay on Opus 5.
- Eulia weekly digest (Sunday 18:00, writes the week's summary). On Opus 5. Writing and synthesis is where Fable 5.1 leads (the professional-work test below), it runs once a week, and nothing irreversible depends on it. Recommendation: the one low-risk place to try Fable 5.1, if you want a live comparison; otherwise stay.
- Staying on Opus 5 regardless: the security red team (on hacking-shaped work Fable 5.1 is served by Opus 4.8), the quarterly QSBS tax evidence job, and the daily immigration timeline review.
Sources
Fable 5.1 & Mythos 5.1 System Card (1 Sep 2026): PDF. Headline table 8.1.A page 167; legal 8.15.2 page 192; professional work 8.15.3 and 8.15.4 page 193; medical 8.17 pages 198 to 199; safety filter 3.2 page 45 and 5.2.1 page 83; injection 5.2.1 to 5.2.2 pages 83 to 89.
Opus 5 System Card (24 Jul 2026): PDF. Legal 8.13.3 page 177; medical 8.15.2 page 185.
Pricing: platform.claude.com pricing page, read 1 Sep 2026. Launch post figures (FrontierFinance, browser agent): anthropic.com.
Cost measurement: token counts summed from the session and sub-agent transcripts on the Mac Studio, 25 Aug to 1 Sep 2026, priced under both rate cards.