Manage costs

Every request Syntic Code makes uses tokens, and tokens are what plans and limits are measured in. This page explains how the CLI estimates cost, where to see it, how to cap it, and the habits that keep a team’s usage predictable.

For what your plan includes and what happens at a limit, see Usage and billing.

How cost is estimated

Each model reply comes back with its token counts: input, output, cache reads and cache writes. Syntic Code turns those into a dollar figure in one of three ways, and labels which one it used:

Cached input is billed at a much lower rate than fresh input, so long sessions that reuse the same context cost less per turn than the raw token count suggests.

Your organization’s rates

If your organization has agreed its own rates, an administrator can put them in the modelPricing key of a managed settings file. Cost figures then use those rates, and the wording changes from a retail estimate to an estimate at your organization’s contract rates.

{
  "modelPricing": {
    "Syntic Code": { "input": 2.0, "output": 10.0, "cache_hit": 0.2 },
    "*": { "multiplier": 0.8 }
  }
}

Model names match regardless of case. An entry that does not parse, or has a negative number, is skipped. modelPricing is only read from managed policy; it has no effect in user or project settings.

See what you are spending

/cost adds up a record for each reply, so it agrees with the footer, survives a resume, and never counts a request twice when you branch a session. Spend copied from the session you branched from is listed as carried over but not counted in the total. Details are in Reading /cost.

Put a ceiling on spend

Scripted runs

For -p runs, especially in CI, set limits so a confused run cannot keep going:

syntic -p "fix the failing test in src/parser" \
  --max-budget-usd 2 --max-turns 30 --output-format json

The budget is checked before each turn, so a run can finish the request in progress and end slightly above the cap. A run stopped by its budget ends with the result subtype error_max_budget_usd; one stopped by its turn limit ends with error_max_turns. Zero, negative or non-numeric budgets are ignored rather than stopping every run. maxTurns in settings sets a default turn limit.

See Headless mode for the rest of the scripting options.

Subagents

Subagents can do a lot of reading. Cap each one:

agents.defaultMaxBudgetUsd can be set by policy, so an organization can cap every subagent on every machine. More in Spend caps.

Spend less per task

Most savings come from sending less context and choosing a lighter model when the task allows.

Context

• Run /clear between unrelated tasks. Old conversation is resent with every request.
• Use /compact in long sessions, or let automatic compaction do it (see Context window).
• Keep SYNTIC.md short and specific. It is loaded into every session. See Memory.
• Turn off MCP servers and plugins you are not using; their tool descriptions take space in every request.
• Point the agent at the right files with @ mentions instead of asking it to search the whole repository.
• Ask subagents to do broad searches. Their reading stays out of your main context.

Model and reasoning

• Use fast mode (/fast) for small edits, renames and quick questions.
• Lower the reasoning effort with /effort for routine work, and raise it only for hard problems.
• Set fallbackModel so a busy model does not stall a long task.

Working style

• Ask for a plan first on large changes. Correcting a plan is cheaper than undoing an implementation.
• Give the checks to run (“run pnpm test when done”) so the agent does not guess.
• Stop a turn that is going the wrong way with Esc instead of letting it finish.

Costs across a team

Syntic Code does not include an organization-wide spend dashboard. To track a team’s usage:

• Send telemetry to your own tools. OpenTelemetry export reports syntic.cost.usage and syntic.token.usage per model, and you can tag every record with team or cost-centre attributes.
• Use your organization’s rates with modelPricing, so the figures match what you pay.
• Set default caps by policy: agents.defaultMaxBudgetUsd for subagents, and maxTurns for runs.
• Limit the model list with availableModels, or set the default with model, so most work uses the model you intend.
• Use usage credits deliberately. If your organization offers credits beyond plan limits, owners and admins control whether they are on in /usage-credits.

A reasonable starting point for a pilot is to let people work freely for two weeks, then look at the exported totals per person and per repository before setting caps. Caps set before you know normal usage tend to interrupt good work.

When the numbers look wrong