AI models & token usage

The Deputy and the Posse call your chosen AI provider with YOUR key — you pay that provider directly, no SubnetSlinger subscription and no markup. Pick your provider in the Deputy: Anthropic (Claude), OpenAI-compatible (ChatGPT, Groq, OpenRouter, Azure OpenAI, LM Studio/vLLM), Google Gemini, or Ollama for a free local model (no cloud, no cost — see "Ollama — run the AI locally"). The app estimates the $ cost for common Anthropic, OpenAI, and Gemini models; an unknown or custom model shows token counts without a $ estimate (check your provider's rates). Cost depends on HOW you use it, lowest to highest: Ollama (free, local) < Haiku 4.5 / GPT-4o-mini / Gemini Flash < Sonnet 4.6 / GPT-4o / Gemini 2.5 Pro < Opus 4.8 < a Posse sweep (highest per run; free on Ollama).

The three models (cheapest → most capable):

• Haiku 4.5 — fastest & cheapest. The DEFAULT. Great for quick lookups and most routine questions.
• Sonnet 4.6 — balanced speed/depth. Good for multi-step investigations.
• Opus 4.8 — deepest reasoning, priciest. Save it for genuinely hard problems.

Pick per chat in the Deputy's model dropdown; it opens on Haiku so you only pay for Opus when you choose it.

How you're billed: per TOKEN, split into input (everything sent — your question + tool results + history) and output (the model's reply). Rough list price, US$ per 1,000,000 tokens — estimates at current pricing (Sept 2026); rates change, so check your provider:

Haiku 4.5 ~ $1 in / $5 out

Sonnet 4.6 ~ $3 in / $15 out

Opus 4.8 ~ $5 in / $25 out

A typical short question is a few thousand tokens — fractions of a cent on Haiku.

The live meter: above the Deputy's composer (and on the Posse screen) an always-on strip shows a Tools on/off pill, 'today ~$X · this chat ~$Y · in N · out N', and 'Ctx N%' (how full the model's context window is) — the same readout the mobile app shows, updating live as the model works, not just at the end. Ollama shows 'local (free)'. The Tools pill is a real switch: OFF sends a plain chat with no tools (cheaper, one round-trip — good for interpreting pasted output); ON lets the Deputy run diagnostics.

Spend caps (three): in AI settings set 'Cap per day' (default $10, ON — a global ceiling across every chat AND the Posse, from a running total that resets at midnight and survives restarts), 'Cap per Posse run' (default $2, ON — bounds one sweep, the priciest action; refuses a launch over the estimate and aborts a sweep that crosses it), and 'Cap this chat' (optional, 0 = off — resets on New chat). All three abort in-flight — on the next turn/round once cost crosses the cap, so a turn already streaming finishes first (bounded to about one turn), not only the next question. A custom model with no published rate is still capped, priced at a safe upper-bound rate so it can't slip past. The STOP button cancels the current request immediately — the emergency brake. Set any cap to 0 to disable it.

Strongest protection — cap the key at your provider: the caps above are enforced by SubnetSlinger. The ceiling nothing can exceed is a hard spend limit on the API key's OWN billing account — set it once and a runaway anywhere (this app, a script, any tool using that key) still can't spend past it. Anthropic: console.anthropic.com → Settings → Limits. OpenAI: platform.openai.com → Settings → Limits (Usage limits). Google: set a budget/quota on the Cloud billing project. Do this first if you're worried about cost; Ollama is local and free, so it needs no cap.

The Posse and your model: the Posse uses the SAME provider and model you set for the Deputy — change it in the Deputy and the Posse follows. On Anthropic it splits into cheaper specialists (Sonnet) + a deep coordinator (Opus); on OpenAI, Gemini, or Ollama every agent uses that backend's one configured model. A sweep is ~6 model passes, so it costs more than one Deputy question (but $0 on Ollama — just heavier on your machine). Run it when you want the thorough pass.

Keeping cost down: SubnetSlinger auto-trims old tool-result payloads from a long conversation (~30k-token cap) so the input doesn't balloon turn after turn; start a New chat to drop context entirely; and stay on Haiku until a problem actually needs Sonnet/Opus. Tool runs themselves (ping, SNMP, etc.) are local and free — you only pay for the model's reasoning over them.

Staying current: the model list and the price table refresh automatically from a hosted manifest (subnetslinger.com/models.json), so new provider models and rate changes appear WITHOUT an app update.