Docs home · Usage · Councils · Cost guide · Orchestration · Architecture eval

Cost Guide

The council spawns real agents, so sessions have real token costs. Two levers control spend:

1. --profile lean|balanced|max — which model tier each spawned agent runs on.
2. /model — which model you (the conductor) run on. The interview, synthesis, design doc, and PRD are conductor work; the engine never overrides your choice.

The engine shows a token estimate at the Assembly approval gate — before anything spawns — and echoes your current conductor model so a mismatch is visible.

What each profile does

Models are routed by spawn site (an agent's model is fixed when it spawns):

Spawn site	lean	balanced (default)	max
Positions & converge (Rounds 1/3)	Sonnet	Opus	Opus
Challenge round (Round 2)	Opus one-shots	persistent agents	Fable one-shots
Audit agents	Sonnet	Sonnet	Opus
Brainstorm	Sonnet	Sonnet	Sonnet
Conductor	your /model	your /model	your /model

Why Round 2 gets special treatment: the challenge round is where positions actually move — it's the highest-leverage spot for extra intelligence. In lean, fresh Opus one-shots read both Round 1 positions from disk and write the rebuttals, so you pay Opus prices only for the debate itself. In max, the same mechanism upgrades to Fable.

Estimated token usage

Total tokens, 5-agent baseline — scale roughly by selected agents ÷ 5:

Mode	lean	balanced	max
brainstorm	10–15K	10–15K	10–15K
quick	20–35K	30–50K	45–75K
standard	80–120K	120–180K	180–270K
deep	+ 50–80K audit	+ 50–100K audit	+ 80–150K audit

Estimates are deliberately coarse — the engine is prompt logic and can't meter tokens. The hard ceiling, where one exists, is the deliberation/audit workflow's token budget (see Orchestration).

If you're API-billed (enterprise agreement, token-conscious)

Run --profile lean. Positions and convergence run on Sonnet 4.6, and Opus is spent only on the challenge round. A Sonnet conductor is acceptable; expect a flatter design doc — if one phase deserves /model opus, it's a standard session's synthesis.

Indicative API pricing (June 2026, per MTok in/out — verify current pricing before budgeting):

Model	Input	Output
Sonnet 4.6	$3.00	$15.00
Opus 4.8	$5.00	$25.00
Fable 5	$10.00	$50.00

Back-of-envelope: a 5-agent standard session at lean lands roughly in the $0.50–1.50 range; balanced roughly $1.50–4.00, output-heavy sessions trending higher.

Other levers:

• Trim the roster at the Assembly gate — cost scales linearly with agent count.
• Prefer --quick for well-understood changes; save standard/deep for genuinely contested designs.
• --auto --profile lean is the cheapest unattended full session.

If you're on a Claude Max plan ($100–200/mo)

There is no per-token dollar cost — sessions draw on your plan's rate limits. The question becomes where to spend quality, not money:

• --profile max puts Opus on every seat and Fable on the challenge round.
• /model opus (or fable) before standard/deep sessions — conductor synthesis is the single highest-leverage model choice in the whole session.
• balanced remains the sane daily default; deep + max burns rate limits fastest, so save it for the decisions that warrant a one-hour deliberation.
• Workflow-backed deliberation also runs in the background, so a max-profile deep session doesn't block your interactive work.

Profile mechanics

• Resolution order: --profile flag → the council's DEFAULT_PROFILE theme variable → balanced.
• The resolved profile persists in session.md/index.json and survives --resume. Pre-v1.2 sessions resume as balanced.
• Composable with every mode: /council --deep --profile lean "Redesign auth".
• If a Fable spawn errors (availability, plan access), the engine retries the spawn with Opus.
• Agent files declare model: inherit; the engine passes the routed tier explicitly on every spawn. Pinned claude-* IDs are rejected by the validator — they go stale.

---

Next: Orchestration — the backends behind a session and how they degrade.