Track token usage and costs

Speedwave monitors model resource consumption through two interfaces: real-time telemetry displayed below the chat composer and an aggregated analytics dashboard tracking historical project usage. Both views read from structured audit records stored at <data_dir>/usage/<project>/proxy/usage.jsonl. See Token and credential paths for storage layouts.

Inspecting project usage

1. Real-time session status: Below the chat composer, inspect input tokens (in), output tokens (out), the conversation cost (chat) and a ring that fills with the context window. Click the ring to open the details: with Anthropic, the context broken down by system prompt, tools, memory files, skills and messages; on an Anthropic subscription, the 5-hour and weekly plan limits with their reset times and any extra usage. A dot on the ring warns when a limit is close or reached. API-key sessions show no plan limits.

2. Terminal statusline: In CLI environments, Claude Code renders equivalent metrics:

   claude-opus-4-8 │ feat/docs │ CTX ██░░░ 40% │ 5h ██░░░ 30% reset 16:42 │ 7d █░░░░ 12% reset 14.04 │ $0.1409

   Rate limits display during Claude subscription sessions, while financial costs display when using direct API keys.

3. Historical analytics dashboard: Open the LLM Usage view to analyze workspace trends.

4. Aggregate summary cards: Inspect request volume, input/output token totals, prompt caching efficiency, decode throughput (tokens/second), error counts, and cumulative cost.

5. Detailed usage breakdowns: Review daily token distribution graphs, provider usage matrices, hourly activity heatmaps, and per-model consumption tables.

Speedwave deduplicates usage events using provider response identifiers. Corrupted records are logged and reported as skipped without interrupting metric aggregation.

Real-time estimation and finalized billing

During stream generation, displayed metrics reflect real-time tokenizer estimates. Once requests complete, the proxy records exact token counts from provider response headers.

Provider pricing models

Cost calculations adapt to the active provider configured in Settings → LLM providers (see LLM providers):

• Anthropic (API Key): Computes financial cost from the bundled Anthropic price list, breaking down prompt input, completion output, cache write, and cache read rates for each model (for example, Sonnet 5 at $2 input and $10 output per million tokens, Opus at $5 and $25, Fable 5.1 at $10 and $50). The 1M-token window uses the same rates.
• Anthropic (OAuth Subscription): Displays a dash (-) indicating flat-rate subscription billing.
• Local model endpoints: Displays a dash (-), reflecting zero per-token cost for self-hosted instances.
• OpenRouter: Evaluates generation costs returned by the OpenRouter API. Initial requests may record a temporary deferred state deferred until generation settlement completes.

Failed requests incurring no upstream provider charges are not billed. For detailed internal billing states, refer to the Cost states reference.