/context opens a dashboard with the per-model breakdown.
What gets tracked
- Prompt tokens (uncached input).
- Completion tokens (model output).
- Cache-write tokens (billed at a higher rate by most providers).
- Cache-read tokens (billed at a discount).
- Subagent tokens tracked separately from the main agent.
- Per-model breakdown when the task router mixes providers.
Providers with built-in pricing
Pricing tables ship for the major providers, updated against their public price lists:| Provider | Notes |
|---|---|
| Anthropic | Claude Opus/Sonnet/Haiku with cache-write and cache-read rates |
| OpenAI | GPT-5.4, GPT-4.1, o3, o4-mini |
| Gemini 2.5 Pro/Flash, Gemini 3 Flash/Pro | |
| DeepSeek | V3.2 (chat and reasoner) |
| Groq | Llama 3.3, Llama 4 Scout, Qwen3, GPT-OSS |
| Mistral | Mistral Large/Medium/Small, Codestral, Magistral, Ministral, Pixtral, Devstral |
| Fireworks | Tier-based pricing (Mixtral, Llama 70B+, DeepSeek) |
| GitHub Copilot | Premium-request multiplier-based estimation |
| OpenRouter | Live pricing from the catalog |
| GitHub Models | Per-token via multipliers |
| Ollama, LM Studio, OpenCode free models | $0.00 |
Why it matters
Two tactics cut cost dramatically:- Mix models. Haiku for spark agents, Sonnet for ember agents, Flash for compaction. A task that would cost 0.05 when the exploration phase routes through Haiku.
- Use caching. Cache reads are 10x cheaper on Anthropic, up to 50% off on Groq/Fireworks. SoulForge structures the system prompt and the Soul Map for maximum cache hits — typical cache-hit rates exceed 60%.
UI
The status bar shows the running total in USD. Compact mode shows tokens plus a dollar figure./context opens the detailed view: per-model usage, cache ratio, subagent spend, and the compaction history.
Use /router to assign cheap models to cheap tasks.