Multi-LLM

Info

Belongs to: Tokenomics

Every time MoAI-ADK starts a session, it makes one decision: “which model will this session run on, and how deeply will it think?” This section covers exactly that model choice — the Claude model lineup, reasoning depth (effort), the inheritance rule by which subagents follow the session, and multi-provider routing that mixes Claude and GLM inside one session.

Tokenomics is the umbrella name for every means of obtaining the same quality of result with fewer tokens. Within it, the work of splitting cost comes in two main branches.

• Model choice (this section) — decides which model, at what depth, and at which provider a session runs. Subagents follow the session’s model and reasoning depth as-is, so choosing one session well is the whole assignment.
• Context saving (the Cost Optimization section) — shrinks the payload handed to the model (context diet) and makes the remaining payload reusable at a cheap rate (prompt caching).

The two branches complement each other. However well you pick a model, an overweight context makes cost unmanageable; however much you shrink the context, a wrong model choice spends expensive reasoning on cheap work. This section owns the first branch — “what does a session run on, and under what conditions.” The whole picture is in the Tokenomics Overview.

What a session chooses is a single {model, effort} pair. That pair applies as-is to every agent call in the session. The two axes answer different questions.

• Model (model) — “which model?” Chooses among Fable, Opus, Sonnet, and Haiku. Changing the model changes both the per-token price and the context window.
• effort (reasoning depth) — “how deeply should it think?” Within the same model, decides whether to skim shallowly or dig deep. Deeper reasoning costs more tokens.

Because the two axes move together, you can run an “expensive model shallowly” or a “cheap model deeply.” That is why effort alone can find the knee of the cost curve without changing the model class.

Info

The lineup and the choice are different things. The table above only shows “available models.” MoAI-ADK’s session default follows the No-Haiku policy, taking the Opus class as the first-line model, and Haiku appears nowhere in the default combination. What actually runs is decided by the session’s model choice — the rule is covered in The Session Picks the Model and Effort below.

effort comes in five levels.

xhigh and max are supported on Opus 5.5, Opus 5, Opus 4.8, Sonnet 5, and Opus 4.7. The shortcut that turns both on at once is the ultrathink keyword. It sets effort: xhigh and at the same time enables Adaptive Thinking (letting the model allocate reasoning tokens on its own).

Warning

Fixed reasoning budgets are forbidden. Opus 4.7 and later reject a fixed reasoning budget such as budget_tokens. Always control reasoning depth through the effort level and Adaptive Thinking — hard-coding a fixed value fails the request.

effort can be changed with a slash command.

/effort low       # speed first
/effort high      # deep reasoning
/effort xhigh     # high difficulty
/effort ultracode # toggle for automatic workflow orchestration
/effort auto      # the model picks based on context

The old profile-matrix approach of picking a model for each agent one by one has retired. Since v3.2, subagents follow the main session’s model and reasoning depth as-is — you pass neither model nor effort when calling a subagent, and MoAI agent definitions declare neither. What remains once the per-agent assignment table is gone is the session’s three choices.

How the old matrix (13 agents × 3 profiles) retired, and the details of the inheritance rule, are covered in Profile Matrix and Model Policy.

flowchart TD
    A["Session model · effort
/model · /effort · profile default"] --> B["Main session runs"]
    B --> C["Subagent call
no model · effort args"]
    C --> D["Runs with the session's model · effort"]
    D --> E{"Provider decided by run mode"}
    E -->|"moai cc"| F["Claude API only"]
    E -->|"moai glm"| G["GLM API only
z.ai backend"]
    F --> I["Agent execution"]
    G --> I

    style A fill:#cc785c,color:#fff
    style I fill:#059669,color:#fff

The last question of model assignment is “which provider runs it.” Beyond the Claude API, MoAI-ADK uses z.ai GLM (Generative Language Model) as an alternative backend. No code changes are needed — it stays Claude-Code-compatible and runs as-is once the environment variables change.

Switching to GLM assigns a matching GLM model to each Claude tier. The pairs injected through Claude Code’s ANTHROPIC_DEFAULT_*_MODEL environment variables are:

glm-5.3-flash is the default model. glm-5.3 remains selectable in any tier slot — name the slot in llm.yaml (llm.glm.models.*) and it loads with the existing behavior (1M context, standard effort collapse) unchanged.

Choose the Claude or GLM launcher explicitly.

moai cg has been retired. It exits with a migration diagnostic without starting Claude or GLM. It is not an alias for moai cc. Projects with llm.team_mode: cg must make an explicit migration choice before launching a session. CG retirement and migration

# 1. Save your GLM API key (once)
moai glm setup sk-your-glm-api-key

# 2. Pick a mode
moai cc            # Claude only
moai glm           # GLM only

Model policy today is the act of setting the session’s model and reasoning depth. Subagents follow the session’s model and reasoning depth as-is, so calls need no explicit model and agent definitions declare neither. In the past, drift — a gap between the “declared model” and the “actually resolved model” — was caught from audit logs; now that inheritance is the default, an explicit model on a call is itself the rare observation worth noting.

What the session model policy decides and does not, and the history of how the old matrix retired, unfold in Model Policy.

• CG retirement and migration
• Model Policy — session model policy, effort fallback, inheritance rule details
• Profile Matrix — where the old matrix retired, and today’s inheritance rule

• Cost Optimization — the other branch of tokenomics: context diet and prompt caching
• Tokenomics Overview — the whole picture joining model assignment and context saving