claude-fable-5-1 appears to over-provision parallel agents in self-authored Workflow scripts vs claude-fable-5 — 25 agents vs 15-limit default, ~10x cost on an equivalent step

Summary

Across a series of same-project, same-type tasks, switching the default model from claude-fable-5 to claude-fable-5-1 correlates with a large jump in how many agents the model chooses to spawn when it authors its own Workflow tool scripts — well past the tool's own stated default ("medium — keep workflows under 15 agents").

Evidence (3 comparable tasks, same pipeline, same kind of "research" workflow step)

| Task | Workflow step | Model | Agents spawned |
|---|---|---|---|
| Task A | research | claude-fable-5 | 6 |
| Task A | review-panel | claude-fable-5 | 3 |
| Task B | research | claude-fable-5 | 7 |
| Task B | review-panel | claude-fable-5 | 3 |
| Task B | visual-audit | claude-fable-5 | 1 |
| Task C | research | claude-fable-5-1 | 19 recorded / 25 actual transcripts |
| Task C | review-panel | claude-fable-5-1 | 3 |
| Task C | visual-audit | claude-fable-5-1 | 1 |

Every Task A/B workflow — all on claude-fable-5 — stayed at or under 7 agents, comfortably inside the tool's stated default cap. The one Task C workflow on claude-fable-5-1 doing the equivalent "research" step jumped to 19-25 agents, blowing past the 15-agent default guideline with no explicit request from the user for a larger scale.

Impact

• That one over-provisioned workflow run alone consumed ~78.8M "effective" tokens (~196M raw: input + cache-read + cache-creation + output) — vs. ~4M tokens for the comparable 7-agent research step on claude-fable-5.
• Several sub-agent tasks inside that run failed mid-execution with You've hit your session limit · resets 6:40pm (Asia/Tokyo) — tokens were spent without producing complete output, and the limit was hit because of this same workflow's size.
• End to end, this made an otherwise-comparable task cost ~3x a directly comparable prior one on claude-fable-5, almost entirely attributable to this single workflow's agent count.

Details

• Session ID: 305fea7e-d1e0-4020-a85c-d0974dcacbaf
• Workflow run ID: wf_92276ab1-05f (defaultModel: "claude-fable-5-1")
• Comparable prior-task session IDs: 64afcfdf-ee2e-4432-9094-e86d22c5b5c4 (claude-fable-5), 258488dc-4dc3-40a2-acd9-3f7be43e9a67 (claude-fable-5)

Ask

Please check whether claude-fable-5-1 has a different tendency than claude-fable-5 when it authors its own multi-agent Workflow scripts — specifically whether it's under-weighting the tool's stated "keep under 15 agents" default guideline when deciding how many parallel research/review agents to spawn.