Cost and tokens

Cost and tokens

Every turn in Ptah is costed in real time. You see tokens in, tokens out, cache reads, and USD cost per message — and a running total for the session — without having to leave the chat.

Where costs appear

Location	What it shows
Cost bar (chat header)	Running session totals: input tokens, output tokens, cache reads, total USD.
Per-message footer	Tokens and USD for that single turn, including all sub-agents spawned from it.
Execution Tree node	Tokens and USD for each individual sub-agent and tool call.
Session summary card	End-of-session rollup with breakdowns per provider and per model.

Where pricing comes from

• Claude (direct Anthropic API) — published Anthropic pricing, per model.
• OpenRouter — live pricing pulled from the OpenRouter model registry. Prices update whenever OpenRouter updates theirs, so third-party models (Gemini, GPT, Moonshot, Z-AI, etc.) are costed accurately without app updates.
• Copilot — reported as $0 because billing is handled by your Copilot subscription. Tokens are still counted.
• Codex / OpenAI direct — official OpenAI pricing per model.
• Ollama (local and cloud) — reported as $0 for local; Ollama Cloud uses their published rates.
• Ptah CLI — reported as $0 (billing is handled by whatever provider the CLI wraps).

Note

If a provider returns an exact cost in its response (some APIs do), Ptah uses the reported value. Otherwise it multiplies token counts by the per-token price from the registry.

Cache tokens

Claude and several OpenRouter-backed models support prompt caching. Cached reads are dramatically cheaper than fresh input tokens, and Ptah shows them as a separate line item:

• Cache creation — first-time tokens that get written to the cache.
• Cache read — tokens served from the cache (≈ 10% of the fresh price on Anthropic).

Seeing a high cache-read count is a good sign: it means Ptah is reusing context across turns.

Per-session rollup

When you close a chat (or click Session summary in the chat header), you get a breakdown:

• Total USD, grouped by provider and model.
• Tokens in, tokens out, and cached tokens.
• Sub-agent spawn count and total sub-agent cost.
• Longest running tool call and slowest turn.

This is the same data written to your local usage logs for cost tracking and analytics.

Tip

If costs look wrong, open the Execution Tree for the suspect turn. Every node carries its own token and USD numbers — usually the culprit is a sub-agent that pulled in a large folder attachment.

MCP chip

Next to the cost bar, an MCP chip shows how many MCP servers this session can reach. It reads a plain connected count normally, and turns amber with a <connected>/<total> count when a server needs authorization or has failed. Click it for a popover listing every server and status, with an Authorize button on any that need one.

See Connectors for the full breakdown, including what the chip does when your claude.ai connectors are unavailable.