Operations

Plans and ceilings

Every tenant resolves to one of four tiers, and the tier decides four things about every turn they take.

The table

Defaults, configurable under AI:Tiers:

Tier	Chat / analysis model	Agent + vision model	Output ceiling	Tool turns	Deep analysis
Free	`gpt-5.6-luna`	`gpt-4o-mini`	8,000	3	no
Starter	`gpt-5.6-luna`	`gpt-4o-mini`	12,000	5	no
Growth	`gpt-5.6-luna`	`gpt-4o-mini`	16,000	6	yes
Enterprise	`gpt-5.6-luna`	`gpt-4o`	25,000	8	yes

Note what does not change across the first three rows: the conversational model is the same on every tier. What a higher tier buys is budget and capability — a longer answer, more tool rounds, and deep analysis — rather than a better model for ordinary questions.

The stronger model is reserved for the agent and vision paths on Enterprise, where the reasoning matters most and the turn count is lowest.

Resolving a tenant to a tier

The tenant record carries plan_code, and the code decides the tier. The three plan codes the product issues come first; the older prefix rule is what is left.

`plan_code`	Tier	Why
`free`	Free	Development and testing only; a signup tenant is always here
`bring_your_ai`	Free	Tether calls no model on this door, so it gets the smallest model budget rather than one it is not paying for. It buys governance, not managed models
`use_tether_ai`	Starter	This door buys managed models

Any other value is read by its prefix, case-insensitively:

`plan_code` starts with	Tier
`ENTERPRISE`	Enterprise
`GROWTH`	Growth
`STARTER`	Starter
`FREE`	Free
anything else, or empty	Free

An unrecognised plan code resolves to the cheapest budget rather than the most generous one. Resolution is cached for 10 minutes, so a plan change takes up to that long to take effect. An operator assigns the plan when issuing a key; see Signup.

What each ceiling stops

Output tokens bound one response. Reasoning tokens come out of the same budget — that is why these are 8k–25k rather than the 1–2k an answer of this length suggests.

Tool turns bound the model-calls-tools loop. Reaching the ceiling is not an error — the model answers with what it gathered, and telemetry records the tool count so you can see which questions hit the wall.

Deep analysis gates capabilityMode: "Deep". A Deep request from a tenant on Free or Starter is silently clamped to Standard rather than rejected, and any tool registered deep_analysis_only is simply not offered.

Ceilings that are not per tier

Bound	Default	Setting
Concurrent tool calls per turn	4	`AI:Orchestration:MaxConcurrentTools`
Per-attempt tool timeout	20s	`ToolTransport:TimeoutSeconds`
Tool response size	256 KB	`ToolTransport:MaxResponseBytes`
Request timeout	60s	`AI:Unified:TimeoutSeconds`
Conversation history	10 messages / 7 days	`AI:Policy:HistoryWindowDays`
Deep-analysis tool turns	12, clamped to the tier	`AI:DeepOps:MaxToolTurns`

Deep analysis raises the tool-turn ceiling but is still clamped by the tier: a Growth tenant requesting Deep gets 6 rounds, not 12.

Changing the table

{
  "AI": {
    "Tiers": {
      "Growth": {
        "ChatModel": "gpt-5.6-luna",
        "AnalysisModel": "gpt-5.6-luna",
        "AgentModel": "gpt-4o-mini",
        "VisionModel": "gpt-4o-mini",
        "MaxOutputTokens": 16000,
        "MaxToolTurns": 6,
        "DeepOpsAllowed": true
      }
    }
  }
}

The table is stated explicitly in configuration so a model or ceiling change is a config edit rather than a code change.

Every model named here must have a price entry, or the process will not start. See Configuration.

A known piece of debt

Plans do not gate tools

A tier controls the model, the ceilings and deep analysis. It does not control which tools a tenant has: that is their catalog, and deep_analysis_only is the only visibility flag a plan influences.