Compare LLM API prices before selecting a model

dsh-model-pricing is a DeepSeek Harness (DSH) plugin that does two jobs inside the harness: it helps you compare LLM API prices before you pick a coding-agent model, and it shows what your sessions actually cost after you used it.

Before choosing, it puts a model pricing and capability table on DSH's Settings -> Models page, covering every connectable model across 213 providers (~7,250 priced models): per-1M-token input, output, and cache-read price, context window, max output, and capability tags (Coding, Agentic, Vision, Long context, Open weights, Structured output) computed from structured catalog fields, with user-configurable tagging rules. For models DSH can actually invoke, the harness's bundled pi-ai catalog overrides the public catalog as the authoritative price, and divergences larger than 10% are flagged row by row. Cross-provider comparison is deliberately worded: the winner badge says "lowest listed", a claim about the catalog rather than the market; $0 free tiers compare only against other free tiers, and a route listed below half the first-party price is flagged to verify. A community promotion feed renders verified, time-boxed offers as badges with provenance tooltips and a one-click filter; offers arrive through reviewed pull requests and are validated by CI.

After using a model, the session costs panel replays the harness's own local session logs, including mid-session model switches, and sums what each project spent per provider and model, priced by the route actually called. Confidence is explicit (routed, then listed, then an estimated range, otherwise nothing is shown), and promo-adjusted savings appear next to list price. Cache-leak attribution shows what share of spend is history re-billed at full input price instead of the cached price, split by cause: idle gaps past documented provider cache TTLs, route switches, and unattributed events; compaction is never counted as a leak. A route performance quadrant places each provider and model by time-to-first-token and tokens per second, taken as medians over your own history and paired with list price, and nothing is labelled until at least five samples exist. The UI is available in English, Chinese, and Russian, with parity checked in CI.

The session panel, cache-leak attribution, and performance quadrant shipped in releases 0.3.0 through 0.5.0. Still planned, not available yet: badges inside provider cards, a /pricing command, and cost projection before compaction.

Use it inside DSH when picking or switching a coding-agent model and you want price, context window, and capability comparisons filtered to routes you can call, or when you want to see which projects cost the most and how much your prompt cache is losing. Do not use it to reconcile a bill: prices, costs, and leak figures are estimates for model selection, and your provider's invoice is the only pricing authority.

Inputs are public catalog data (fetched on first load and cached locally), the harness's reachable model routes, and the local DSH session logs under your DSH home directory. Outputs are the pricing table with badges and flags on Settings -> Models, plus the session costs panel with per-workspace spend, cache leaks by cause, and the price versus speed quadrant.

Install with:

dsh plugin --profile web add dsh-model-pricing

Then restart dsh web and open Settings -> Models; the table appears at the bottom of the page and no configuration is required.

Runs entirely on your machine with no server of its own. Outbound traffic is an anonymous read of public data from models.dev (213 providers), the @earendil-works/pi-ai catalog already bundled with DSH, and the promotion feed served as static JSON from the plugin repository. Session costs and cache leaks are computed from local logs and are never uploaded. Wrong prices or missing promotions can be reported through structured issue forms or a pull request to promos/**, where every entry needs a source URL and an expiry. Listed in awesome-dsh-plugin, accepted 2026-09-11.

DSH users choosing coding-agent models who want price, context, and capability comparisons in the harness, and who want to understand their real per-project spend and cache losses without sending logs anywhere. Apache-2.0 licensed.